Learning Joint Spatial-Temporal

Transformations for Video Inpainting

Yanhong Zeng1,2∗ , Jianlong Fu3† , and Hongyang Chao1,2†

1
School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China
2
Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of
Education, China
arXiv:2007.10247v1 [cs.CV] 20 Jul 2020

3
Microsoft Research Asia
[email protected], [email protected], [email protected]

Abstract. High-quality video inpainting that completes missing regions

in video frames is a promising yet challenging task. State-of-the-art ap-
proaches adopt attention models to complete a frame by searching miss-
ing contents from reference frames, and further complete whole videos
frame by frame. However, these approaches can suffer from inconsis-
tent attention results along spatial and temporal dimensions, which often
leads to blurriness and temporal artifacts in videos. In this paper, we pro-
pose to learn a joint Spatial-Temporal Transformer Network (STTN)
for video inpainting. Specifically, we simultaneously fill missing regions
in all input frames by self-attention, and propose to optimize STTN by a
spatial-temporal adversarial loss. To show the superiority of the proposed
model, we conduct both quantitative and qualitative evaluations by us-
ing standard stationary masks and more realistic moving object masks.
Demo videos are available at https://github.com/researchmm/STTN.

Keywords: Video Inpainting; Generative Adversarial Networks

1 Introduction
Video inpainting is a task that aims at filling missing regions in video frames with
plausible contents [2]. An effective video inpainting algorithm has a wide range of
practical applications, such as corrupted video restoration [10], unwanted object
removal [22,26], video retargeting [16] and under/over-exposed image restoration
[18]. Despite of the huge benefits of this technology, high-quality video inpainting
still meets grand challenges, such as the lack of high-level understanding of videos
[15,29] and high computational complexity [5,33].
Significant progress has been made by using 3D convolutions and recurrent
networks for video inpainting [5,16,29]. These approaches usually fill missing
regions by aggregating information from nearby frames. However, they suffer
from temporal artifacts due to limited temporal receptive fields. To solve the
above challenge, state-of-the-art methods apply attention modules to capture
∗
This work was done when Y. Zeng was an intern at Microsoft Research Asia.
†
J. Fu and H. Chao are the corresponding authors.
2 Y. Zeng, J. Fu, and H. Chao

our results input+mask

Fig. 1. We propose Spatial-Temporal Transformer Networks for completing missing

regions in videos in a spatially and temporally coherent manner. The top row shows
sample frames with yellow masks denoting user-selected regions to be removed. The
bottom row shows our completion results. [Best viewed with zoom-in]

long-range correspondences, so that visible contents from distant frames can be

used to fill missing regions in a target frame [18,25]. One of these approaches
synthesizes missing contents by a weighting sum over the aligned frames with
frame-wise attention [18]. The other approach proposes a step-by-step fashion,
which gradually fills missing regions with similar pixels from boundary towards
the inside by pixel-wise attention [25]. Although promising results have been
shown, these methods have two major limitations due to the significant appear-
ance changes caused by complex motions in videos. One limitation is that these
methods usually assume global affine transformations or homogeneous motions,
which makes them hard to model complex motions and often leads to inconsis-
tent matching in each frame or in each step. Another limitation is that all videos
are processed frame by frame without specially-designed optimizations for tem-
poral coherence. Although post-processing is usually used to stabilize generated
videos, it is usually time-costing. Moreover, the post-processing may fail in cases
with heavy artifacts.
To relieve the above limitations, we propose to learn a joint Spatial-Temporal
Transformer Network (STTN) for video inpainting. We formulate video in-
painting as a “multi-to-multi” problem, which takes both neighboring and dis-
tant frames as input and simultaneously fills missing regions in all input frames.
To fill missing regions in each frame, the transformer searches coherent contents
from all the frames along both spatial and temporal dimensions by a proposed
multi-scale patch-based attention module. Specifically, patches of different scales
are extracted from all the frames to cover different appearance changes caused
by complex motions. Different heads of the transformer calculate similarities on
spatial patches across different scales. Through such a design, the most relevant
patches can be detected and transformed for the missing regions by aggregating
attention results from different heads. Moreover, the spatial-temporal transform-
ers can be fully exploited by stacking multiple layers, so that attention results
for missing regions can be improved based on updated region features. Last but
not least, we further leverage a spatial-temporal adversarial loss for joint opti-
 Spatial-Temporal Transformer Networks for Video Inpainting 3

mization [5,6]. Such a loss design can optimize STTN to learn both perceptually
pleasing and coherent visual contents for video inpainting.
In summary, our main contribution is to learn joint spatial and temporal
transformations for video inpainting, by a deep generative model with adver-
sarial training along spatial-temporal dimensions. Furthermore, the proposed
multi-scale patch-based video frame representations can enable fast training and
inference, which is important to video understanding tasks. We conduct both
quantitative and qualitative evaluations using both stationary masks and mov-
ing object masks for simulating real-world applications (e.g., watermark removal
and object removal). Experiments show that our model outperforms the state-
of-the-arts by a significant margin in terms of PSNR and VFID with relative
improvements of 2.4% and 19.7%, respectively. We also show extensive ablation
studies to verify the effectiveness of the proposed spatial-temporal transformer.

2 Related Work

To develop high-quality video inpainting technology, many efforts have been

made on filling missing regions with spatially and temporally coherent contents
in videos [2,13,18,24,29,33]. We discuss representative patch-based methods and
deep generative models for video inpainting as below.
Patch-based methods: Early video inpainting methods mainly formulate
the inpainting process as a patch-based optimization problem [1,7,26,31]. Specif-
ically, these methods synthesize missing contents by sampling similar spatial
or spatial-temporal patches from known regions based on a global optimization
[24,27,31]. Some approaches try to improve performance by providing foreground
and background segments [10,26]. Other works focus on joint estimations for
both appearance and optical-flow [13,22]. Although promising results can be
achieved, patch-based optimization algorithms typically assume a homogeneous
motion field in holes and they are often limited by complex motion in general
situations. Moreover, optimization-based inpainting methods often suffer from
high computational complexity, which is infeasible for real-time applications [33].
Deep generative models: With the development of deep generative mod-
els, significant progress has been made by deep video inpainting models. Wang
et al. are the first to propose to combine 3D and 2D fully convolution networks
for learning temporal information and spatial details for video inpainting [29].
However, the results are blurry in complex scenes. Xu et al. improve the perfor-
mance by jointly estimating both appearance and optical-flow [33,37]. Kim et
al. adopt recurrent networks for ensuring temporal coherence [16]. Chang et al.
develop Temporal SN-PatchGAN [35] and temporal shift modules [19] for free-
form video inpainting [5]. Although these methods can aggregate information
from nearby frames, they fail to capture visible contents from distant frames.
To effectively model long-range correspondences, recent models have adopted
attention modules and show promising results in image and video synthesis
[21,34,36]. Specifically, Lee et al. propose to synthesize missing contents by
weighted summing aligned frames with frame-wise attention [18]. However, the
4 Y. Zeng, J. Fu, and H. Chao

Spatial-Temporal Transformer multi-layers

patch scale n:
frame 1 patch scale 1: frame 1
Q

1x1
softmax

3x3
target frame K target frame

1x1 attention

V
frame t frame t
1x1

input frames extract patches batch dot output frames

with frame- with frame-
level encoder level decoder

Fig. 2. Overview of the Spatial-Temporal Transformer Networks (STTN).

STTN consists of 1) a frame-level encoder, 2) multi-layer multi-head spatial-temporal
transformers and 3) a frame-level decoder. The transformers are designed to simulta-
neously fill holes in all input frames with coherent contents. Specifically, a transformer
matches the queries (Q) and keys (K) on spatial patches across different scales in mul-
tiple heads, thus the values (V) of relevant regions can be detected and transformed
for the holes. Moreover, the transformers can be fully exploited by stacking multiple
layers to improve attention results based on updated region features. 1 × 1 and 3 × 3
denote the kernel size of 2D convolutions. More details can be found in Section 3.

frame-wise attention relies on global affine transformations between frames,

which is hard to handle complex motions. Oh et al. gradually fill holes step
by step with pixel-wise attention [25]. Despite promising results, it is hard to en-
sure consistent attention result in each recursion. Moreover, existing deep video
inpainting models that adopt attention modules process videos frame by frame
without specially-designed optimization for ensuring temporal coherence.

3 Spatial-Temporal Transformer Networks

3.1 Overall design

Problem formulation: Let X1T := {X1 , X2 , ..., XT } be a corrupted video se-

quence of height H, width W and frames length T . M1T := {M1 , M2 , ..., MT }
denotes the corresponding frame-wise masks. For each mask Mi , value “0” indi-
cates known pixels, and value “1” indicates missing regions. We formulate deep
video inpainting as a self-supervised task that randomly creates (X1T , M1T ) pairs
as input and reconstruct the original video frames Y1T = {Y1 , Y2 , ..., YT }. Specif-
ically, we propose to learn a mapping function from masked video X1T to the
output Ŷ1T := {Ŷ1 , Ŷ2 , ..., ŶT }, such that the conditional distribution of the real
data p(Y1T |X1T ) can be approximated by the one of generated data p(Ŷ1T |X1T ).
 Spatial-Temporal Transformer Networks for Video Inpainting 5

The intuition is that an occluded region in a current frame would probably

be revealed in a region from a distant frame, especially when a mask is large
or moving slowly. To fill missing regions in a target frame, it is more effective
to borrow useful contents from the whole video by taking both neighboring
frames and distant frames as conditions. To simultaneously complete all the
input frames in a single feed-forward process, we formulate the video inpainting
task as a “multi-to-multi” problem. Based on the Markov assumption [11], we
simplify the “multi-to-multi” problem and denote it as:
T
Y
t+n t+n
p(Ŷ1T |X1T ) = p(Ŷt−n |Xt−n T
, X1,s ), (1)
t=1

t+n
where Xt−n denotes a short clip of neighboring frames with a center moment t
T
and a temporal radius n. X1,s denotes distant frames that are uniformly sampled
T T
from the videos X1 in a sampling rate of s. Since X1,s can usually cover most
key frames of the video, it is able to describe “the whole story” of the video.
Under this formulation, video inpainting models are required to not only preserve
temporal consistency in neighboring frames, but also make the completed frames
to be coherent with “the whole story” of the video.
Network design: The overview of the proposed Spatial-Temporal Transfo-
rmer Networks (STTN) is shown in Figure 2. As indicated in Eq. (1), STTN
t+n T
takes both neighboring frames Xt−n and distant frames X1,s as conditions, and
complete all the input frames simultaneously. Specifically, STTN consists of three
components, including a frame-level encoder, multi-layer multi-head spatial-
temporal transformers, and a frame-level decoder. The frame-level encoder is
built by stacking several 2D convolution layers with strides, which aims at encod-
ing deep features from low-level pixels for each frame. Similarly, the frame-level
decoder is designed to decode features back to frames. Spatial-temporal trans-
formers are the core component, which aims at learning joint spatial-temporal
transformations for all missing regions in the deep encoding space.

3.2 Spatial-temporal transformer

To fill missing regions in each frame, spatial-temporal transformers are designed

to search coherent contents from all the input frames. Specifically, we propose
to search by a multi-head patch-based attention module along both spatial and
temporal dimensions. Different heads of a transformer calculate attentions on
spatial patches across different scales. Such a design allows us to handle appear-
ance changes caused by complex motions. For example, on one hand, attentions
for patches of large sizes (e.g., frame size H × W ) aim at completing station-
ary backgrounds. On the other hand, attentions for patches of small sizes (e.g.,
H W
10 × 10 ) encourage capturing deep correspondences in any locations of videos
for moving foregrounds.
A multi-head transformer runs multiple “Embedding-Matching-Attending”
steps for different patch sizes in parallel. In the Embedding step, features of each
6 Y. Zeng, J. Fu, and H. Chao

frame are mapped into query and memory (i.e., key-value pair) for further re-
trieval. In the Matching step, region affinities are calculated by matching queries
and keys among spatial patches that are extracted from all the frames. Finally,
relevant regions are detected and transformed for missing regions in each frame
in the Attending step. We introduce more details of each step as below.
Embedding: We use f1T = {f1 , f2 , ..., fT }, where fi ∈ Rh×w×c to denote the
features encoded from the frame-level encoder or former transformers, which is
the input of transformers in Fig. 2. Similar to many sequence modeling models,
mapping features into key and memory embeddings is an important step in
transformers [9,28]. Such a step enables modeling deep correspondences for each
region in different semantic spaces:
qi , (ki , vi ) = Mq (fi ), (Mk (fi ), Mv (fi )), (2)
where 1 ≤ i ≤ T , Mq (·), Mk (·) and Mv (·) denote the 1 × 1 2D convolutions that
embed input features into query and memory (i.e., key-value pair) feature spaces
while maintaining the spatial size of features.
Matching: We conduct patch-based matching in each head. In practice, we
first extract spatial patches of shape r1 × r2 × c from the query feature of each
frame, and we obtain N = T × h/r1 × w/r2 patches. Similar operations are con-
ducted to extract patches in the memory (i.e., key-value pair in the transformer).
Such an effective multi-scale patch-based video frame representation can avoid
redundant patch matching and enable fast training and inference. Specifically, we
reshape the query patches and key patches into 1-dimension vectors separately,
so that patch-wise similarities can be calculated by matrix multiplication. The
similarity between i-th patch and j-th patch is denoted as:
pqi · (pkj )T
si,j = √ , (3)
r1 × r2 × c
where 1 ≤ i, j ≤ N , pqi denotes the i-th query patch, pkj denotes the j-th key
patch. The similarity value is normalized by the dimension of each vector to avoid
a small gradient caused by subsequent softmax function [28]. Corresponding
attention weights for all patches are calculated by a softmax function:

PN
exp(si,j )/ exp(si,n ), pj ∈ Ω,


αi,j = n=1 (4)


0, pj ∈ Ω̄.


where Ω denotes visible regions outside masks, and Ω̄ denotes missing regions.
Naturally, we only borrow features from visible regions for filling holes.
Attending: After modeling the deep correspondences for all spatial patches,
the output for the query of each patch can be obtained by weighted summation
of values from relevant patches:
N
X
oi = αi,j pvj , (5)
j=1
 Spatial-Temporal Transformer Networks for Video Inpainting 7

target frame t=1 t=5

output frame t = 15 t = 20

Fig. 3. Illustration of the attention maps for missing regions learned by

STTN. For completing the dog corrupted by a random mask in a target frame (e.g.,
t=10), our model is able to “track” the moving dog over the video in both spatial and
temporal dimensions. Attention regions are highlighted in bright yellow.

where pvj denotes the j-th value patch. After receiving the output for all patches,
we piece all patches together and reshape them into T frames with original spatial
size h × w × c. The resultant features from different heads are concatenated and
further passed through a subsequent 2D residual block [12]. This subsequent
processing is used to enhance the attention results by looking at the context
within the frame itself.
The power of the proposed transformer can be fully exploited by stacking
multiple layers, so that attention results for missing regions can be improved
based on updated region features in a single feed-forward process. Such a multi-
layer design promotes learning coherent spatial-temporal transformations for
filling in missing regions. As shown in Fig. 3, we highlight the attention maps
learned by STTN in the last layer in bright yellow. For the dog partially occluded
by a random mask in a target frame, spatial-temporal transformers are able to
“track” the moving dog over the video in both spatial and temporal dimensions
and fill missing regions in the dog with coherent contents.

3.3 Optimization objectives

As outlined in Section 3.1, we optimize the proposed STTN in an end-to-end
manner by taking the original video frames as ground truths without any other
labels. The principle of choosing optimization objectives is to ensure per-pixel
reconstruction accuracy, perceptual rationality and spatial-temporal coherence
in generated videos [5,8,14,18]. To this end, we select a pixel-wise reconstruction
loss and a spatial-temporal adversarial loss as our optimization objectives.
In particular, we include L1 losses calculated between generated frames and
original frames for ensuring per-pixel reconstruction accuracy in results. The L1
8 Y. Zeng, J. Fu, and H. Chao

losses for hole regions are denoted as:

kM1T  (Y1T − Ŷ1T )k1

Lhole = , (6)
kM1T k1

and corresponding L1 losses for valid regions are denoted as:

k(1 − M1T )  (Y1T − Ŷ1T )k1

Lvalid = , (7)
k1 − M1T k1

where  indicates element-wise multiplication, and the values are normalized by

the size of corresponding regions.
Inspired by the recent studies that adversarial training can help to ensure
high-quality content generation results, we propose to use a Temporal Patch-
GAN (T-PatchGAN) as our discriminator [5,6,34,36]. Such an adversarial loss
has shown promising results in enhancing both perceptual quality and spatial-
temporal coherence in video inpainting [5,6]. In particular, the T-PatchGAN is
composed of six layers of 3D convolution layers. The T-PatchGAN learns to dis-
tinguish each spatial-temporal feature as real or fake, so that spatial-temporal
coherence and local-global perceptual details of real data can be modeled by
STTN. The detailed optimization function for the T-PatchGAN discriminator
is shown as follows:

LD = Ex∼PY T (x) [ReLU (1 − D(x))] + Ez∼PŶ T (z) [ReLU (1 + D(z))], (8)

1 1

and the adversarial loss for STTN is denoted as:

Ladv = −Ez∼PŶ T (z) [D(z)]. (9)

1

The overall optimization objectives are concluded as below:

L = λhole · Lhole + λvalid · Lvalid + λadv · Ladv . (10)

We empirically set the weights for different losses as: λhole = 1, Lvalid = 1,
Ladv = 0.01. Since our model simultaneously complete all the input frames in a
single feed-forward process, our model runs at 24.3 fps on a single GPU NVIDIA
V100. More details are provided in the Section D of our supplementary material.

4 Experiments
4.1 Dataset
To evaluate the proposed model and make fair comparisons with SOTA ap-
proaches, we adopt the two most commonly-used datasets in video inpainting,
including Youtube-VOS [32] and DAVIS [3]. In particular, YouTube-VOS con-
tains 4,453 videos with various scenes, including bedrooms, streets, and so on.
The average video length in Youtube-VOS is about 150 frames. We follow the
 Spatial-Temporal Transformer Networks for Video Inpainting 9

original train/validation/test split (i.e., 3,471/474/508) and report experimen-

tal results on the test set for Youtube-VOS. In addition, we also evaluate dif-
ferent approaches on DAVIS dataset [3], as this dataset is composed of 150
high-quality videos of challenging camera motions and foreground motions. We
follow the setting in previous works [16,33], and set the training/testing split
as 60/90 videos. Since the training set of DAVIS is limited (60 videos with at
most 90 frames for each), we initialize model weights by a pre-trained model on
YouTube-VOS following the settings used in [16,33].
To simulate real-world applications, we evaluate models by using two types
of free-form masks, including stationary masks and moving masks [6,16,18]. Be-
cause free-form masks are closer to real masks and have been proved to be
effective for training and evaluating inpainting models [5,6,20,23]. Specifically,
for testing stationary masks, we generate stationary random shapes as testing
masks to simulate applications like watermark removal. More details of the gen-
eration algorithm are provided in the Section B of our supplementary material.
Since this type of application targets at reconstructing original videos, we take
original videos as ground truths and evaluate models from both quantitative and
qualitative aspects. For testing moving masks, we use foreground object anno-
tations as testing masks to simulate applications like object removal. Since the
ground truths after foreground removal are unavailable, we evaluate the models
through qualitative analysis following previous works [16,18,33].

4.2 Baselines and evaluation metrics

Recent deep video inpainting approaches have shown state-of-the-art perfor-
mance with fast computational time [16,18,25,33]. To evaluate our model and
make fair comparisons, we select the most recent and the most competitive ap-
proaches for comparisons, which are listed as below:
– VINet [16] adopts a recurrent network to aggregate temporal features from
neighboring frames.
– DFVI [33] fills missing regions in videos by pixel propagation algorithm
based on completed optical flows.
– LGTSM [6] proposes a learnable temporal shift module and a spatial-
temporal adversarial loss for ensuring spatial and temporal coherence.
– CAP [18] synthesizes missing contents by a deep alignment network and a
frame-based attention module.
We fine-tune baselines multiple times on YouTube-VOS [32] and DAVIS [3] by
their released models and codes and report their best results in this paper.
We report quantitative results by four numeric metrics, i.e., PSNR [33], SSIM
[5], flow warping error [17] and video-based Fréchet Inception Distance (VFID)
[5,30]. Specifically, we use PSNR and SSIM as they are the most widely-used
metrics for video quality assessment. Besides, the flow warping error is included
to measure the temporal stability of generated videos. Moreover, FID has been
proved to be an effective perceptual metric and it has been used by many inpaint-
ing models [25,30,38]. In practice, we use an I3D [4] pre-trained video recognition
model to calculate VFID following the settings in [5,30].
10 Y. Zeng, J. Fu, and H. Chao

4.3 Comparisons with state-of-the-arts

Quantitative Evaluation: We report quantitative results for filling stationary

masks on Youtube-VOS [32] and DAVIS [3] in Table 1. As stationary masks
often involve partially occluded foreground objects, it is challenging to recon-
struct a video especially with complex appearances and object motions. Table
1 shows that, compared with SOTA models, our model performs better video
reconstruction quality with both per-pixel and overall perceptual measurements.
Specifically, our model outperforms the SOTA models by a significant margin,
especially in terms of PSNR, flow warp error and VFID. The specific gains are
2.4%, 1.3% and 19.7% relative improvements on Youtube-VOS, respectively. The
superior results show the effectiveness of the proposed spatial-temporal trans-
former and adversarial optimizations in STTN.

Models PSNR? SSIM (%)? Ewarp (%)† VFID†

Youtube-vos

VINet[16] 29.20 94.34 0.1490 0.072

DFVI [33] 29.16 94.29 0.1509 0.066
LGTSM [6] 29.74 95.04 0.1859 0.070
CAP [18] 31.58 96.07 0.1470 0.071
Ours 32.34 96.55 0.1451 0.053
VINet[16] 28.96 94.11 0.1785 0.199
DAVIS

DFVI [33] 28.81 94.04 0.1880 0.187

LGTSM [6] 28.57 94.09 0.2566 0.170
CAP [18] 30.28 95.21 0.1824 0.182
Ours 30.67 95.60 0.1779 0.149

Table 1. Quantitative comparisons with state-of-the-art models on Youtube-VOS [32]

and DAVIS [3]. Our model outperforms baselines in terms of PSNR [33], SSIM [5], flow
warping error (Ewarp ) [17] and VFID [30]. ? Higher is better. † Lower is better.

Qualitative Evaluation: For each video from test sets, we take all frames
for testing. To compare visual results from different models, we follow the setting
used by most video inpainting works and randomly sample three frames from
the video for case study [18,25,29]. We select the most three competitive models,
DFVI [33], LGTSM [6] and CAP [18] for comparing results for stationary masks
in Fig. 4. We also show a case for filling in moving masks in Fig. 5. To conduct
pair-wise comparisons and analysis in Fig. 5, we select the most competitive
model, CAP [18], according to the quantitative comparison results. We can find
from the visual results that our model is able to generate perceptually pleasing
and coherent contents in results. More video cases are available online§ .
In addition to visual comparisons, we visualize the attention maps learned
by STTN in Fig. 6. Specifically, we highlight the top three relevant regions
captured by the last transformer in STTN in bright yellow. The relevant regions
§
video demo: https://github.com/researchmm/STTN
 Spatial-Temporal Transformer Networks for Video Inpainting 11

input frame DFVI LGTSM CAP ours

Fig. 4. Visual results for stationary masks. The first column shows input frames
from DAVIS [3] (top-3) and YouTube-VOS [32] (bottom-3), followed by results from
DFVI [33], LGTSM [6], CAP [18], and our model. Comparing with the SOTAs, our
model generates more coherent structures and details of the legs and boats in results.

are selected according to the attention weights calculated by Eq. (4). We can find
in Fig. 6 that STTN is able to precisely attend to the objects for filling partially
occluded objects in the first and the third cases. For filling the backgrounds in
the second and the fourth cases, STTN can correctly attend to the backgrounds.
User Study: We conduct a user study for a more comprehensive comparison.
we choose LGTSM [6] and CAP [18] as two strong baselines, since we have
observed their significantly better performance than other baselines from both
quantitative and qualitative results. We randomly sampled 10 videos (5 from
DAVIS and 5 from YouTube-VOS) for stationary masks filling, and 10 videos
from DAVIS for moving masks filling. In practice, 28 volunteers are invited to
the user study. In each trial, inpainting results from different models are shown
to the volunteers, and the volunteers are required to rank the inpainting results.
To ensure a reliable subjective evaluation, videos can be replayed multiple times
by volunteers. Each participant is required to finish 20 groups of trials without
time limit. Most participants can finish the task within 30 minutes. The results
of the user study are concluded in Fig 7. We can find that our model performs
better in most cases for these two types of masks.
12 Y. Zeng, J. Fu, and H. Chao

input frame CAP ours

Fig. 5. Visual comparisons for filling moving masks. Comparing with CAP [18],
one of the most competitive models for filling moving masks, our model is able to
generate visually pleasing results even under complex scenes (e.g., clear faces for the
first and the third frames, and better results than CAP for the second frame).

t = 25 t = 25 t = 12 t = 14 t = 30

t = 43 t = 43 t = 40 t = 50 t = 60

t = 40 t = 40 t=8 t = 16 t = 28

t = 30 t = 30 t = 34 t = 46 t = 52
input frame output frame attention map

Fig. 6. Illustration of attention maps for missing regions learned by the

proposed STTN. We highlight the most relevant patches in yellow according to
attention weights. For filling partially occluded objects (the first and the third cases),
STTN can precisely attend to the objects. For filling backgrounds (the second and the
fourth cases), STTN can correctly attend to the backgrounds.
 Spatial-Temporal Transformer Networks for Video Inpainting 13

(a) stationary holes (b) moving holes

60% 60%

40% 40%

20% 20%

0% 0%
rank 1 rank 2 rank 3 rank 1 rank 2 rank 3
Ours CAP LGTSM Ours CAP LGTSM

Fig. 7. User study. “Rank x” means the percentage of results from each model being
chosen as the x-th best. Our model is ranked in first place in most cases.

4.4 Ablation Study

To verify the effectiveness of the spatial-temporal transformers, this section

presents ablation studies on DAVIS dataset [3] with stationary masks. More
ablation studies can be found in the Section E of our supplementary material.
Effectiveness of multi-scale: To verify the effectiveness of using multi-
scale patches in multiple heads, we compare our model with several single-head
STTNs with different patch sizes. In practice, we select patch sizes according
to the spatial size of features, so that the features can be divided into patches
without overlapping. The spatial size of features in our experiments is 108 × 60.
Results in Table 2 show that our full model with multi-scale patch-based video
frame representation achieves the best performance under this setting.

Patch size PSNR? SSIM(%)? Ewarp (%)† VFID†

108 × 60 30.16 95.16 0.2243 0.168
36 × 20 30.11 95.13 0.2051 0.160
18 × 10 30.17 95.20 0.1961 0.159
9×5 30.43 95.39 0.1808 0.163
Ours 30.67 95.60 0.1779 0.149

Table 2. Ablation study by using different patch scales in attention layers. Ours com-
bines the above four scales. ? Higher is better. † Lower is better.

Effectiveness of multi-layer: The spatial-temporal transformers can be

stacked by multiple layers to repeat the inpainting process based on updated
region features. We verify the effectiveness of using multi-layer spatial-temporal
transformers in Table 3. We find that stacking more transformers can bring
continuous improvements and the best results can be achieved by stacking eight
layers. Therefore, we use eight layers in transformers as our full model.
14 Y. Zeng, J. Fu, and H. Chao

Stack PSNR? SSIM(%)? Ewarp (%)† VFID†

×2 30.17 95.17 0.1843 0.162
×4 30.38 95.37 0.1802 0.159
×6 30.53 95.47 0.1797 0.155
×8 (ours) 30.67 95.60 0.1779 0.149

Table 3. Ablation study by using different stacking number of the proposed spatial-
temporal transformers. ? Higher is better. † Lower is better.
our results input+mask

Fig. 8. A failure case. The bottom row shows our results with enlarged patches in the
bottom right corner. For reconstructing the dancing woman occluded by a large mask,
STTN fails to generate continuous motions and it generates blurs inside the mask.

5 Conclusions
In this paper, we propose a novel joint spatial-temporal transformation learn-
ing for video inpainting. Extensive experiments have shown the effectiveness
of multi-scale patch-based video frame representation in deep video inpainting
models. Coupled with a spatial-temporal adversarial loss, our model can be op-
timized to simultaneously complete all the input frames in an efficient way. The
results on YouTube-VOS [32] and DAVIS [3] with challenging free-form masks
show the state-of-the-art performance by our model.
We note that STTN may generate blurs in large missing masks if continuous
quick motions occur. As shown in Fig. 8, STTN fails to generate continuous
dancing motions and it generates blurs when reconstructing the dancing woman
in the first frame. We infer that STTN only calculates attention among spatial
patches, and the short-term temporal continuity of complex motions are hard to
capture without 3D representations. In the future, we plan to extend the pro-
posed transformer by using attention on 3D spatial-temporal patches to improve
the short-term coherence. We also plan to investigate other types of temporal
losses [17,30] for joint optimization in the future.

Acknowledgments
This project was supported by NSF of China under Grant 61672548, U1611461.
 Spatial-Temporal Transformer Networks for Video Inpainting 15

References
1. Barnes, C., Shechtman, E., Finkelstein, A., Goldman, D.B.: Patchmatch: A ran-
domized correspondence algorithm for structural image editing. TOG 28(3), 24:1–
24:11 (2009)
2. Bertalmio, M., Bertozzi, A.L., Sapiro, G.: Navier-stokes, fluid dynamics, and image
and video inpainting. In: CVPR. pp. 355–362 (2001)
3. Caelles, S., Montes, A., Maninis, K.K., Chen, Y., Van Gool, L., Perazzi, F., Pont-
Tuset, J.: The 2018 davis challenge on video object segmentation. arXiv (2018)
4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the
kinetics dataset. In: CVPR. pp. 6299–6308 (2017)
5. Chang, Y.L., Liu, Z.Y., Lee, K.Y., Hsu, W.: Free-form video inpainting with 3d
gated convolution and temporal patchgan. In: ICCV. pp. 9066–9075 (2019)
6. Chang, Y.L., Liu, Z.Y., Lee, K.Y., Hsu, W.: Learnable gated temporal shift module
for deep video inpainting. In: BMVC (2019)
7. Criminisi, A., Pérez, P., Toyama, K.: Region filling and object removal by exemplar-
based image inpainting. TIP 13(9), 1200–1212 (2004)
8. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional
neural networks. In: CVPR. pp. 2414–2423 (2016)
9. Girdhar, R., Carreira, J., Doersch, C., Zisserman, A.: Video action transformer
network. In: CVPR. pp. 244–253 (2019)
10. Granados, M., Tompkin, J., Kim, K., Grau, O., Kautz, J., Theobalt, C.: How not to
be seenobject removal from videos of crowded scenes. Computer Graphics Forum
31(21), 219–228 (2012)
11. Hausman, D.M., Woodward, J.: Independence, invariance and the causal markov
condition. The British journal for the philosophy of science 50(4), 521–583 (1999)
12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: CVPR. pp. 770–778 (2016)
13. Huang, J.B., Kang, S.B., Ahuja, N., Kopf, J.: Temporally coherent completion of
dynamic video. TOG 35(6), 1–11 (2016)
14. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and
super-resolution. In: ECCV. pp. 694–711 (2016)
15. Kim, D., Woo, S., Lee, J.Y., Kweon, I.S.: Deep blind video decaptioning by tem-
poral aggregation and recurrence. In: CVPR. pp. 4263–4272 (2019)
16. Kim, D., Woo, S., Lee, J.Y., Kweon, I.S.: Deep video inpainting. In: CVPR. pp.
5792–5801 (2019)
17. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning
blind video temporal consistency. In: ECCV. pp. 170–185 (2018)
18. Lee, S., Oh, S.W., Won, D., Kim, S.J.: Copy-and-paste networks for deep video
inpainting. In: ICCV. pp. 4413–4421 (2019)
19. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video under-
standing. In: ICCV. pp. 7083–7093 (2019)
20. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpaint-
ing for irregular holes using partial convolutions. In: ECCV. pp. 85–100 (2018)
21. Ma, S., Fu, J., Wen Chen, C., Mei, T.: Da-gan: Instance-level image translation by
deep attention generative adversarial networks. In: CVPR. pp. 5657–5666 (2018)
22. Matsushita, Y., Ofek, E., Ge, W., Tang, X., Shum, H.Y.: Full-frame video stabi-
lization with motion inpainting. TPAMI 28(7), 1150–1163 (2006)
23. Nazeri, K., Ng, E., Joseph, T., Qureshi, F., Ebrahimi, M.: Edgeconnect: Generative
image inpainting with adversarial edge learning. In: ICCVW (2019)
16 Y. Zeng, J. Fu, and H. Chao

24. Newson, A., Almansa, A., Fradet, M., Gousseau, Y., Pérez, P.: Video inpainting
of complex scenes. SIAM Journal on Imaging Sciences 7(4), 1993–2019 (2014)
25. Oh, S.W., Lee, S., Lee, J.Y., Kim, S.J.: Onion-peel networks for deep video com-
pletion. In: ICCV. pp. 4403–4412 (2019)
26. Patwardhan, K.A., Sapiro, G., Bertalmio, M.: Video inpainting of occluding and
occluded objects. In: ICIP. pp. 11–69 (2005)
27. Patwardhan, K.A., Sapiro, G., Bertalmı́o, M.: Video inpainting under constrained
camera motion. TIP 16(2), 545–553 (2007)
28. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser,
L., Polosukhin, I.: Attention is all you need. In: NeurIPS. pp. 5998–6008 (2017)
29. Wang, C., Huang, H., Han, X., Wang, J.: Video inpainting by jointly learning
temporal structure and spatial details. In: AAAI. pp. 5232–5239 (2019)
30. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.:
Video-to-video synthesis. In: NeuraIPS. pp. 1152–1164 (2018)
31. Wexler, Y., Shechtman, E., Irani, M.: Space-time completion of video. TPAMI
29(3), 463–476 (2007)
32. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos:
A large-scale video object segmentation benchmark. arXiv (2018)
33. Xu, R., Li, X., Zhou, B., Loy, C.C.: Deep flow-guided video inpainting. In: CVPR.
pp. 3723–3732 (2019)
34. Yang, F., Yang, H., Fu, J., Lu, H., Guo, B.: Learning texture transformer network
for image super-resolution. In: CVPR. pp. 5791–5800 (2020)
35. Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., Huang, T.S.: Free-form image inpainting
with gated convolution. In: ICCV. pp. 4471–4480 (2019)
36. Zeng, Y., Fu, J., Chao, H., Guo, B.: Learning pyramid-context encoder network
for high-quality image inpainting. In: CVPR. pp. 1486–1494 (2019)
37. Zhang, H., Mai, L., Xu, N., Wang, Z., Collomosse, J., Jin, H.: An internal learning
approach to video inpainting. In: CVPR. pp. 2720–2729 (2019)
38. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable
effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)
 Spatial-Temporal Transformer Networks for Video Inpainting 17

Supplementary Material

This supplementary material presents the details of complete video inpainting

results in Section A and our stationary mask generation algorithm in Section B.
We provide the details of our network architectures in Section C and the imple-
mentation details in Section D. Finally, extensive ablation studies and analysis
for the proposed Spatial-Temporal Transformer Networks for video inpainting
can be found in Section E.

A Video Inpainting Results

To compare visual results from different inpainting models in our main paper,
we follow the setting used in most video inpainting works [13,16,33]. Specifically,
we sample several frames from video results and show them in Figure 4 and
Figure 5 in the main paper. However, sampled frames cannot truly reflect video
results. Sometimes sampled static frames look less blurry but artifacts can be
stronger in a dynamic video. Therefore, we provide 20 video cases for a more
comprehensive comparison¶ .
In practice, we test all the videos in the test sets of DAVIS dataset [3] (90
cases) and Youtube-VOS dataset [32] (508 cases), and we randomly show 20 cases
for visual comparisons. Specifically, five cases from DAVIS and five cases from
Youtube-VOS are used to test filling stationary masks. Since Youtube-VOS has
no dense object annotations, we sample 10 videos with dense object annotations
from DAVIS to test filling moving masks following the setting used in previous
works [16,18,33]. To conduct side-by-side comparisons and analysis, we select
the two most competitive video inpainting models, LGTSM [6] and CAP [18] in
the videos. LGTSM and CAP are fine-tuned multiple times to achieve optimal
video results by the codes and models publicly provided by their official Github
homepagek . We can find from the video results that our model outperforms the
state-of-the-art models in most cases.

B Stationary Mask Generation Algorithm

Inspired by Xu et al. [33], we use stationary masks and moving masks as testing
masks to simulate real-world applications (e.g., watermark removal and object
removal) in the main paper. As introduced in Section 4.1 in the main paper, on
one hand, we use frame-wise foreground object annotations from DAVIS datasets
[3] as moving masks to simulate applications like object removal. On the other
hand, we generate random shapes as stationary masks to simulate applications
like watermark removal. Specifically, for the task of removing watermarks, a user
often draw a mask along the outline of a watermark. Inspired by previous mask
¶
video demo: https://github.com/researchmm/STTN
k
LGTSM: https://github.com/amjltc295/Free-Form-Video-Inpainting
CAP: https://github.com/shleecs/Copy-and-Paste-Networks-for-Deep-Video-Inpainting
18 Y. Zeng, J. Fu, and H. Chao

generation algorithms [5,35], we propose a stationary mask generation algorithm

to simulate such a behavior for drawing masks for watermarks. Specifically, the
proposed algorithm randomly generates a set of control points around a unit
circle, and then it smoothly connects these points into a closed cyclic contour
by cubic Bezier curves. The details of the stationary mask generation algorithm
are shown in Algorithm 1 as follows.

Algorithm 1 Algorithm for stationary mask generation. maxPointNum,

maxLength are hyper-parameters to control the statinary mask generation.
mask = zeros(imgHeight, imgWidth)
pointNum = random.uniform(maxPointNum)
startX = origX = random.uniform(imgWidth)
startY = origY = random.uniform(imgHeight)
angles = linspace(0, 2*pi, pointNum)
for i=0 to pointNum do
length = random.uniform(maxLength)
x = sin(angles[i]) * length
y = cos(angles[i]) * length
// comment: ensuring smoothness of contours
Connect (startX, startY) to (x, y) by cubic Bezier curves.
startX = x
startY = y
end for
// comment: ensuring a closed cyclic contour
Connect (startX, startY) to (origX, origY) by cubic Bezier curves.

C Details of Network Architecture

The Spatial-Temporal Transformer Network (STTN) is built upon a generative

adversarial framework. Specifically, the proposed STTN plays a role as a gener-
ator in the framework, and we adopt a Temporal PatchGAN (T-PatchGAN) [5]
as our discriminator. The T-PatchGAN is composed of six layers of 3D convolu-
tion layers. Specifically, the T-PatchGAN learns to classify each spatial-temporal
feature as real or fake, while STTN learns to fool the T-PatchGAN. Such an ad-
versarial training allows STTN to model the local-global perceptual rationality
and the spatial-temporal coherence of real videos [5]. In addition to the introduc-
tion in Section 3 in the main paper, we provide the details of the architectures
of STTN and the T-PatchGAN in Table 4 and Table 5, respectively. Specifi-
cally, features inside holes are computed by dilated 2D convolutions. We argue
that STTN is able to leverages multi-scale contexts and updates holes’ features
multiple times to improve attention results.
 Spatial-Temporal Transformer Networks for Video Inpainting 19

Module Name Filter Size # Channels Stride/Up Factor Nonlinearity

2dConv 3×3 64 2 LeakyReLU(0.2)
2dConv 3×3 64 1 LeakyReLU(0.2)
2dConv 3×3 128 2 LeakyReLU(0.2)
2dConv 3×3 256 1 LeakyReLU(0.2)
1×1 1 -
Transformer × 8 256
3×3 1 LeakyReLU(0.2)
BilinearUpSample - 256 2 -
2dConv 3×3 128 1 LeakyReLU(0.2)
2dConv 3×3 64 1 LeakyReLU(0.2)
BilinearUpSample - 64 2 -
2dConv 3×3 64 1 LeakyReLU(0.2)
2dConv 3×3 3 1 Tanh

Table 4. Details of the proposed Spatial-Temporal Transformer Networks (STTN).

“2dConv” means 2D convolution layers. “Transformer × 8” denotes stacking the pro-
posed spatial-temporal transformers by eight layers. A transformer layer involves 1 × 1
and 3 × 3 convolutions (The overview of STTN is shown in Fig. 2 in the main paper).
We use bilinear interpolations for all upsample operations on feature maps [20,25]. We
show whether and what nonlinearity layer is used in the nonlinearity column.

Module Name Filter Size # Channels Stride Nonlinearity

SN-3dConv 3×5×5 64 (1,2,2) LeakyReLU(0.2)
SN-3dConv 3×5×5 128 (1,2,2) LeakyReLU(0.2)
SN-3dConv 3×5×5 256 (1,2,2) LeakyReLU(0.2)
SN-3dConv 3×5×5 256 (1,2,2) LeakyReLU(0.2)
SN-3dConv 3×5×5 256 (1,2,2) LeakyReLU(0.2)
SN-3dConv 3×5×5 256 (1,2,2) -

Table 5. Details of the Temporal-PatchGAN (T-PatchGAN) discriminator [5]. The

T-PatchGAN is composed of six 3D convolution layers. “SN-3dConv” denotes a 3D
convolution layer that adopts spectral normalization to stabilize GAN’s training [5].
20 Y. Zeng, J. Fu, and H. Chao

D Implementation details
Hyper-parameters: To maintain the aspect ratio of videos and take into ac-
count the memory limitations of modern GPUs, we resize all video frames into
432 × 240 for both training and testing [13,16,18,33]. During training, we set the
batch size as 8, and the learning rate starts with 1e-4 and decays with factor
0.1 every 150k iterations. Specifically, for each iteration, we sample five frames
from a video in a consecutive or discontinuous manner with equal probability
for training following Lee et al. [18,25].
Computation complexity: Our full model has a total of 12.6M train-
able parameters. It costs about 3.9G GPU memory for completing a video from
DAVIS dataset [3] by STTN on average. The proposed multi-scale patch-based
video frame representations can enable fast training and inference. Specifically,
our model runs at about 24.3fps with an NVIDIA V100 GPU and it runs at
about 10.43 fps with an NVIDIA P100 GPU on average. Its total training time
was about 3 days on YouTube-VOS dataset [32] and one day for fine-tuning on
DAVIS dataset [3] with 8 Tesla V100 GPUs. The computation complexity of the
proposed spatial-temporal transformers are denoted as:
D  
X HW 2
O( 2 · (n · ) · (pw ph Cl ) + nkl2 HW Cl−1 Cl ) ≈ O(n2 ), (11)
pw ph
l=1

where D is the number of transformer layers, n is the number of input frames,

HW is the feature size, pw ph is the patch size, kl denotes for kernel size, and
C is the channel number of features. In Eq. (11), we focus on the computa-
tion complexity caused by the spatial-temporal transformers and leave out other
computation costs (e.g., encoding and decoding costs) for simplification.

E More ablation studies

To verify the effectiveness of the proposed Spatial-Temporal Transformer Net-
works (STTN) for video inpainting, this section presents extensive ablation stud-
ies on DAVIS dataset [3] with stationary masks.
Effectiveness of utilizing distant frames: we test our full model with dif-
ferent sample rates to prove the benefits of utilizing distant frames. Quantitative
comparison results on DAVIS dataset [3] with stationary masks can be found
in Table 6. The first row (s > T ) means that the STTN takes only neighboring
frames as input. Besides, the second row (s = 20) means that the STTN takes
both neighboring frames and distant frames that are uniformly sampled from
the videos in a sampling rate of 20 frames.
Table 6 shows that leveraging visible contexts in distant frames helps in gen-
erating better results especially in terms of VFID with 5.70% relative improve-
ments. Based on the observation that most videos in YouTube-VOS dataset [32]
and DAVIS dataset [3] won’t vary a lot within 10 frames on average, we set the
sample rate as 10 in our full model to avoid sampling redundant frames and to
save computation costs.
 Spatial-Temporal Transformer Networks for Video Inpainting 21

Sample Rate PSNR? SSIM(%)? Ewarp (%)† VFID†

s>T 30.55 95.47 0.1802 0.158
s = 20 30.62 95.55 0.1790 0.152
s = 10 (ours) 30.67 95.60 0.1779 0.149

Table 6. Ablation study by utilizing distant frames in different sampling rates. Our
full model set s = 10. ? Higher is better. † Lower is better.

Effectiveness of masked normalization: As shown in Eq. (3) and Eq.

(4) in the main paper, we normalize the value of similarity by the dimension
of vectors and filter out unknown regions for similarities calculating. In this
part, we conduct comparisons between models with or without such a masked
normalization in Table 7. Results show that such an operation is necessary since
it brings improvements with a significant margin comparing with the one without
masked normalization.

PSNR? SSIM(%)? Ewarp (%)† VFID†

w/o masked norm. 30.39 95.32 0.1849 0.162
w/ masked norm. 30.67 95.60 0.1779 0.149

Table 7. Ablation study for the effectiveness of masked normalization operation on

similarity calculation. ? Higher is better. † Lower is better.

Effectiveness of the Temporal PatchGAN Loss: Recent state-of-the-

art deep video inpainting models that adopt attention modules often include a
perceptual loss [14] and a style loss [8] as optimization objectives for perceptually
pleasing results [18,25]. However, they do not leverage specially-designed losses
for ensuring temporal coherence. Chang et al. propose a novel Temporal Patch-
GAN (T-PatchGAN) loss for ensuring both perceptual rationality and spatial-
temporal coherence of videos [5,6]. However, they only apply T-PatchGAN on
consecutive frames while the attention-based deep video inpainting models take
discontinuous frames as input for training. We are the first to introduce T-
PatchGAN in video inpainting models that adopt attention modules and show
that T-PatchGAN is also powerful in discontinuous frames. Such a joint opti-
mization encourages STTN to learn both local-global perceptual rationality and
coherent spatial-temporal transformations for video inpainting.
We verify the effectiveness of the T-PatchGAN loss by quantitative compar-
isons in Table 8. Compared with the STTN optimized by a style loss [8] and a
perceptual loss [14] following previous works [18,25], the STTN optimized by a
T-PatchGAN loss performs better by a significant margin, especially in terms
of VFID with 6.9% relative improvements. We also provide a visual comparison
in Fig. 9. The visual results show that the STTN optimized by a T-PatchGAN
22 Y. Zeng, J. Fu, and H. Chao

loss can generate more coherent results than the one optimized by a perceptual
loss and a style loss. The superior results show the effectiveness of the joint
spatial-temporal adversarial learning in STTN.

losses PSNR? SSIM(%)? Ewarp (%)† VFID†

w/ style [8], w/ perceptual [14] 30.38 95.35 0.1821 0.160
w/ T-PatchGAN [5] 30.67 95.60 0.1779 0.149

? †
Table 8. Ablation study for different losses. Higher is better. Lower is better.

input frame w/ perceptual and style losses w/ T-PatchGAN loss

Fig. 9. Visual comparisons between an STTN optimized by a perceptual loss [14] and
a style loss [8] and an STTN optimized by a T-PatchGAN loss [5]. These two models
perform similarly in small missing regions, while in large missing regions, the model
optimized by perceptual and style losses tends to generate artifacts in the missing
regions. [Best viewed with zoom-in]

Specifically, perceptual loss and style loss have shown great impacts in many
image generation tasks since they were proposed [8,14,20]. A perceptual loss
computes L1 distance between the activation maps of real frames and generated
frames. A style loss is similar to the perceptual loss but aims at minimizing the
L1 distance between Gram matrices of the activation maps of real frames and
generated frames. In practice, the activation maps are extracted from layers (e.g.,
 Spatial-Temporal Transformer Networks for Video Inpainting 23

pool1, pool2 and pool3) of a pre-trained classification network (more details see
[18,20,25]). With the help of extracted low-level features, the perceptual loss and
the style loss are helpful in generating high-frequency details.
Unfortunately, perceptual and style losses are calculated on the features of
a single frame and they are unable to leverage temporal contexts. When filling
in a large missing region in videos, the perceptual and style losses are hard to
enforce the generator to synthesize rational contents due to limited contexts.
As a result, they have to generate meaningless high-frequency textures to match
ground truths’ low-level features. For example, for filling the large missing regions
in the second and the third frames in Fig. 9, the STTN optimized by perceptual
and style losses tends to generate high-frequency artifacts in the large missing
regions. Similar artifacts can be found in the failure cases of previous works [5,20].
Since the T-PatchGAN is able to leverage temporal contexts to optimize the
generator, there are fewer artifacts in the results by using the T-PatchGAN. For
the above considerations, we use the T-PatchGAN loss instead of the perceptual
and style losses in our final optimization objectives. In the future, we plan to
design video-based perceptual and style losses which are computed on spatial-
temporal features to leverage temporal contexts for optimization.