IRNet Label Generation Scripts v1

CAMs are generated to create proper .npy files to use as an input in PuzzleCAM repository. (https://github.com/cenkbircanoglu/PuzzleCAM)

Scripts used to execute

sh scripts/irnet/make_cam.sh

ResNet101-v2

Training and Evaluations are done on the Pascal VOC train dataset.

miou = 0.5913 epoch = 249

Experiment logs
https://tensorboard.dev/experiment/axANxgoBREuDSnxSrmsVgg/

Scripts used to execute experiments

sh scripts/train_aug/train.sh

sh scripts/train_aug/make_cam.sh

sh scripts/train_aug/eval.sh

Checkpoints DIR D:\checkpoints\resnet101v2

ResNet50-v2

Training and Evaluations are done on the Pascal VOC train dataset.

miou = 0.5642 epoch = 224

Experiment logs
https://tensorboard.dev/experiment/ubRFjgmqT1qSaHPVflZoIQ/

Scripts used to execute experiments

sh scripts/train_aug/train.sh

sh scripts/train_aug/make_cam.sh

sh scripts/train_aug/eval.sh

Checkpoints DIR D:\checkpoints\resnet50v2

Compare CAM Eval Thresholds

Training and Evaluations are done on the Pascal VOC train dataset.

Experiment logs
https://tensorboard.dev/experiment/T4uqc542SDuwaoQlb7EwZw/

Scripts used to execute experiments

sh scripts/compare_threshold/train.sh

sh scripts/compare_threshold/make_cam.sh

sh scripts/compare_threshold/eval.sh

Checkpoints DIR D:\checkpoints\threshold

Longer Run on Training Dataset

Training and Evaluations are done on the Pascal VOC train dataset.

Best Result is achieved with

backbone=resnet50v2 threshold=0.15 crop_size=512 epoch=249

miou=53.52

Experiment logs
https://tensorboard.dev/experiment/qo1oZeuGTpO01Cjfxknj5g/

Scripts used to execute experiments

sh scripts/long_run/train.sh

sh scripts/long_run/make_cam.sh

sh scripts/long_run/eval.sh

Checkpoints DIR D:\checkpoints\long_run

DeepLab Dilation Rates

Training and Evaluations are done on the Pascal VOC train dataset.

It seems dilation version 1 performs slightly better.

Experiment logs
https://tensorboard.dev/experiment/dWDxvN47QvqFOh366bKmjw/

Scripts used to execute experiments

sh scripts/compare_dilate_version/train.sh

sh scripts/compare_dilate_version/make_cam.sh

sh scripts/compare_dilate_version/eval.sh

Checkpoints DIR D:\checkpoints\dilation

Compare Backbones and Scales v1

Training and Evaluations are done on the Pascal VOC train dataset.

Best Result is achieved with

backbone=resnet50v2 threshold=0.15 crop_size=512

miou=52.66

Experiment logs
https://tensorboard.dev/experiment/j4a3Zn16TqKgokIWAuRm8g/

Scripts used to execute experiments

sh scripts/compare_backbones_train.sh

sh scripts/compare_backbones_make_cam.sh

sh scripts/compare_backbones_eval.sh

sh scripts/compare_scales_eval.sh

Checkpoints DIR D:\checkpoints\backbone_and_scale