Multiple ego-actions and explanations predictions on new annotated self-driving dataset
Architecture: ResNet101 with linear classifier
Usage:
1. Predicting multiple ego-actions:
Training:
python baseline/train_cnn.py --imageroot path/to/image/ --gtroot path/to/groundtruth/actions/ --out_dir output/directory/
Testing:
python baseline/test_cnn.py --imageroot path/to/image/ --gtroot path/to/groundtruth/actions/ --model_root path/to/trained/model/weights
2. Predicting multiple ego-actions and explanations
Training:
python baseline/train_cnn.py --imageroot path/to/image/ --gtroot path/to/groundtruth/actions/ --resonroot path/to/groundtruth/explanations --out_dir output/directory/ --side
Testing:
python baseline/test_cnn.py --imageroot path/to/image/ --gtroot path/to/groundtruth/actions/ --resonroot path/to/groundtruth/explanations --model_root path/to/trained/model/weights --side
Architecture: FasterRCNN with selector and predictor
cd maskrcnn-benchmark/
1. Preparation:
Follow maskrcnn-benchmark/INSTALL.md for installation instructions.
2. Pretrained faster-rcnn
Follow maskrcnn-benchmark/README.md for training and testing faster-rcnn.
Using "e2e_faster_rcnn_R_50_C4_1x.yaml" as config file.
3. Predicting multiple ego-actions and explanations
Training:
python action_prediction/train_all.py --imageroot path/to/image/ --gtroot path/to/groundtruth/actions/ --resonroot path/to/groundtruth/explanations --model_root path/to/pretrained/fasterrcnn/weights OUT_DIR output/directory MODEL.SIDE True
Testing:
python action_prediction/test_all.py --imageroot path/to/image/ --gtroot path/to/groundtruth/actions/ --resonroot path/to/groundtruth/explanations --model_root path/to/pretrained/fasterrcnn/weights OUT_DIR output/directory MODEL.SIDE True
Optional paramters: --initLR, --weight_decay, --num_epoch, --batch_size