Skip to content

Latest commit

 

History

History
86 lines (72 loc) · 4.64 KB

README.md

File metadata and controls

86 lines (72 loc) · 4.64 KB

Learning in the Frequency Domain

Our work is based on mmdetection. mmdetection is an open source object detection toolbox based on PyTorch. It is a part of the open-mmlab project developed by Multimedia Laboratory, CUHK.

Prerequisites

Install

  • Please refer to INSTALL.md for installation and dataset preparation.
  • Download pretrained models and extract to work_dirs. The folder structure should look like this:
work_dirs
├── mask_rcnn_r50_fpn_1x_dct_24_wofreeze
│   ├── 20191029_145538.log
│   └── latest.pth
└── mask_rcnn_r50_fpn_1x_dct_64_wofreeze
    ├── 20191029_151515.log
    └── latest.pth

Evaluation for instance segmentation

Run test.py to start testing

Testing the proposed DCT-24 model for instance segmentation

python tools/test.py configs/mask_rcnn_r50_rpn_1x_DCT_static_24_wofreeze.py work_dirs/mask_rcnn_r50_fpn_1x_dct_24_wofreeze/latest.pth --out results.pkl --eval bbox segm

Testing the proposed DCT-64 model for instance segmentation

python tools/test.py configs/mask_rcnn_r50_rpn_1x_DCT_static_64_wofreeze.py work_dirs/mask_rcnn_r50_fpn_1x_dct_64_wofreeze/latest.pth --out results.pkl --eval bbox segm

Results for instance segmentation

Performance of the proposed model - ResNet-50-FPN

Backbone #Channels Size Per Channel bbox
AP AP@0.5 AP@0.75 APS APM APL
ResNet-50-FPN (RGB) 3 800x1333 37.3 59.0 40.2 21.9 40.9 48.1
DCT-24 (ours) 24 200x334 37.7 59.2 40.9 21.7 41.4 49.1
DCT-64 (ours) 64 200x334 38.1 59.6 41.1 22.5 41.6 49.7
Backbone #Channels Size Per Channel mask
AP AP@0.5 AP@0.75 APS APM APL
ResNet-50-FPN (RGB) 3 800x1333 34.2 55.9 36.2 15.8 36.9 50.1
DCT-24 (ours) 24 200x334 34.6 56.1 36.9 16.1 37.4 50.7
DCT-64 (ours) 64 200x334 35.0 56.5 37.4 16.9 37.6 51.6

Instance segmentation examples generated by the DCT-24 model.

example

Evaluation for object detection

Run test.py to start testing

Testing the proposed DCT-24 model for object detection

python tools/test.py configs/faster_rcnn_r50_fpn_1x_static_24_wofreeze.py work_dirs/faster_rcnn_r50_fpn_1x_dct_24_wofreeze/latest.pth --out results.pkl --eval bbox segm

Testing the proposed DCT-64 model for object detection

python tools/test.py configs/faster_rcnn_r50_fpn_1x_static_64_wofreeze.py work_dirs/faster_rcnn_r50_fpn_1x_dct_64_wofreeze/latest.pth --out results.pkl --eval bbox segm

Results for object detection

Performance of the proposed model - ResNet-50-FPN

Backbone #Channels Size Per Channel bbox
AP AP@0.5 AP@0.75 APS APM APL
ResNet-50-FPN (RGB) 3 800x1333 36.4 58.4 39.1 21.5 40.0 46.6
DCT-24 (ours) 24 200x334 37.2 58.8 39.9 21.9 40.7 48.9
DCT-64 (ours) 64 200x334 37.2 58.5 40.6 21.9 40.9 48.3