Thermal images needs to be rescaled for neural network ssd-efficientdet requires mean and std-dev as hyper-parameter for scaling the inputs sample code with computed mean and std-dev for FLIR's dataset is provided in common_preprocessor.py

ssd-efficientdet model overrides should be passed as dictionary while calling backbone neural network

efficientnet_base = efficientnet_model.EfficientNet.from_name(
        model_name=self._efficientnet_version,
        overrides={'rescale_input': False,'input_channels':1})
        

The overrides above along with the pre-processor will generate a single channel compatible object-detection model which can then be used for inference.

Training the models can be a bit more complicated as the default tfrecord decoders are incompatible with single channel data produced by thermal cameras.