[CVPR 2023] Spatial-temporal Concept based Explanation of 3D ConvNets
a 3D extention to https://github.com/amiratag/ACE
Extract concepts and calculate importance score for each single class:
python main.py --target_class crying --source_dir V:/ViSR-explanation --working_dir ./test/ --model_to_run keras-r3d --labels_path ./data/label.txt --bottlenecks average_pooling3d --num_random_exp 80 --max_videos 500 --min_videos 80 --imageshape 16 112 112 --model_path r3d.h5 --batchsize 60
The model_path is the trained model where it is stored. You should first train a model with your own dataset using Keras. Then explain the model with our method.
Our paper has been accepted in CVPR 2023.
Our research on 3D interpretation is still in progress.