dynamo13 / fcam-wsol

Pytorch implementation of F-CAM. Paper: "F-CAM: Full Resolution Class Activation Maps via Guided Parametric Upscaling".

Pytorch 1.9.0 code for:

F-CAM: Full Resolution Class Activation Maps via Guided Parametric Upscaling (https://arxiv. org/abs/2109.07069)

WACV 2022: [Slides] [Video presentation] [Poster]

Citation:

@InProceedings{belharbi2022fcam,
  title={F-CAM: Full Resolution Class Activation Maps via Guided Parametric Upscaling},
  author={Belharbi, S. and Sarraf, A. and Pedersoli, M. and Ben Ayed, I. and McCaffrey, L. and Granger, E.},
  booktitle = {WACV},
  year={2022}
}

Issues:

Please create a github issue.

Content:

• Requirements
• Datasets
• Run code
• Method
• Results

Requirements:

• Python 3.7.10
• Pytorch 1.9.0
• torchvision 0.10.0
• Full dependencies
• Build and install CRF:
  • Install Swig
  • CRF

cd dlib/crf/crfwrapper/bilateralfilter
swig -python -c++ bilateralfilter.i
python setup.py install

Download datasets :

See folds/wsol-done-right-splits/dataset-scripts. For more details, see wsol-done-right repo.

Once you download the datasets, you need to adjust the paths in get_root_wsol_dataset().

Run code :

• WSOL baselines: CAM over CUB using ResNet50:

time python main_wsol.py --task STD_CL \
                         --encoder_name resnet50 \
                         --arch STDClassifier \
                         --opt__name_optimizer sgd \
                         --batch_size 32 \
                         --opt__step_size 15 \
                         --opt__gamma 0.1 \
                         --max_epochs 50 \
                         --freeze_cl False \
                         --support_background True \
                         --method CAM \
                         --spatial_pooling WGAP \
                         --dataset CUB \
                         --box_v2_metric False \
                         --cudaid $cudaid \
                         --debug_subfolder None \
                         --opt__lr 0.0017 \
                         --exp_id 08_19_2021_14_05_20_620912__6229687

• Once you trained a WSOL baseline, copy the best model from the exp folder into the folder ./pretrained. The best model is located in a folder with the form name CUB-resnet50-CAM-WGAP-cp_best-boxv2_False. Copy the whole folder.
• F-CAM: to train with F-CAM, a pretrained WSOL model needs to be prepared as in the previous step. Run for training with F-CAM:

time python main_wsol.py --task F_CL \
                        --encoder_name resnet50 \
                        --arch UnetFCAM \
                        --opt__name_optimizer sgd \
                        --batch_size 32 \
                        --eval_checkpoint_type best \
                        --opt__step_size 1000 \
                        --opt__gamma 0.1 \
                        --max_epochs 50 \
                        --freeze_cl True \
                        --support_background True \
                        --method CAM \
                        --spatial_pooling WGAP \
                        --dataset CUB \
                        --box_v2_metric False \
                        --cudaid $cudaid \
                        --debug_subfolder None \
                        --opt__lr 0.01 \
                        --elb_init_t 1.0 \
                        --elb_max_t 10.0 \
                        --elb_mulcoef 1.01 \
                        --sl_fc True \
                        --sl_fc_lambda 1.0 \
                        --sl_start_ep 0 \
                        --sl_end_ep -1 \
                        --sl_min 1 \
                        --sl_max 1 \
                        --sl_ksz 3 \
                        --sl_min_p 0.1 \
                        --sl_fg_erode_k 11 \
                        --sl_fg_erode_iter 1 \
                        --crf_fc True \
                        --crf_lambda 2e-09 \
                        --crf_sigma_rgb 15.0 \
                        --crf_sigma_xy 100.0 \
                        --crf_scale 1.0 \
                        --crf_start_ep 0 \
                        --crf_end_ep -1 \
                        --max_sizepos_fc True \
                        --max_sizepos_fc_lambda 0.1 \
                        --max_sizepos_fc_start_ep 0 \
                        --max_sizepos_fc_end_ep -1 \
                        --entropy_fc False \
                        --exp_id 08_19_2021_14_09_48_915565__1492324

Method:

Results: