Discovering an inference recipe for weakly-supervised object localization

This is the official PyTorch implementation for "Discovering an inference recipe for weakly-supervised object localization" Pattern Recognition.

Datasets

Our repository enables evaluation and training of various WSOL methods on three benchmarks, CUB and ImageNet, and OpenImages.

Methods

Implemetation of the methods are provided by NAVER wsol evaluation. For more detail, please refer to it.

Run train+eval

We support the following architectures, methods, and ways to genete CAM:

• Architectures.

  • vgg16
  • inception_v3
  • resnet50

• Methods

  • cam
  • has
  • acol
  • spg
  • adl
  • cutmix

• Ways to generate CAMs(neg : nwc, bin : binarize weight to 0 or 1, vote : ours)

  • origin
  • neg
  • bin
  • vote

Below is an example command line for the train+eval script.

python main.py --dataset_name OpenImages \
               --architecture vgg16 \
               --wsol_method cam \
               --experiment_name OpenImages_vgg16_CAM \
               --pretrained TRUE \
               --num_val_sample_per_class 5 \
               --large_feature_map FALSE \
               --batch_size 32 \
               --epochs 10 \
               --lr 0.00227913316 \
               --lr_decay_frequency 3 \
               --weight_decay 5.00E-04 \
               --override_cache FALSE \
               --workers 4 \
               --box_v2_metric True \
               --iou_threshold_list 30 50 70 \
               --eval_checkpoint_type last

Below is an example command line for the eval script. For CUB and ILSVRC, this script evaluate the models using follwing metrics. For fair top1 accuracy, set box_v2_metric as FALSE. Choose the way to generate CAM by setting mode(origin, neg, bin, vote). If you use our method vote, it needs to set binary_rate.

• metrics.
  • top1 accuracy
  • top1 classification
  • gt-known accuracy

python main_top1.py --dataset_name CUB \
               --architecture vgg16 \
               --wsol_method cam \
               --experiment_name cub_vgg16_CAM \
               --pretrained TRUE \
               --num_val_sample_per_class 5 \
               --large_feature_map FALSE \
               --batch_size 32 \
               --epochs 10 \
               --lr 0.00227913316 \
               --lr_decay_frequency 3 \
               --weight_decay 5.00E-04 \
               --override_cache FALSE \
               --workers 4 \
               --box_v2_metric FALSE \
               --iou_threshold_list 30 50 70 \
               --mode vote \
               --binary_rate 0.15 \
               --eval_checkpoint_type last

See config.py or config_1.py for the full descriptions of the arguments, especially the method-specific hyperparameters.