Pixel-wise Gradient Uncertainty for Convolutional Neural Networks applied to Out-of-Distribution Segmentation

The present code was utilized to generate the results in "Pixel-wise Gradient Uncertainty for Convolutional Neural Networks applied to Out-of-Distribution Segmentation".

Method implementation

The implementation of the DeepLabv3+ segmentation model (forked from https://github.com/NVIDIA/semantic-segmentation/tree/sdcnet) can be found in the subfolder segmentation. Gradient norms are computed in the forward pass within segmentation/network/deepv3.py in the classes DeepV3Plus (l.253) and DeepWV3Plus (l.349) as the final chunk of layers is replaced by a child class MySequential (ll.35-124) of nn.Sequential. Norm computation is implemented in the method grad_heatmap.

The forward pass can otherwise be called with minimal alterations (concerning the number of pixel-wise output features) to the scripts segmentation/demo_folder.py or segmentation/eval_gradients.py provided by the original framework by NVIDIA (with the provided weights of the original repository).

Evaluation for pixel-wise and segment-wise UQ

The scripts for pixel- and segment-wise uncertainty evaluations in Sec. 4.1 and 4.2 can be found in the subfolder evaluation which are run by calls within evaluation/main.py. In order to run evaluation, make sure to correctly define the path variables in evaluation/global_defs.py.

Evaluation for OoD segmentation

See official benchmark at http://github.com/SegmentMeIfYouCan/road-anomaly-benchmark.