Gradual Guidance Learning with Middle Feature for Weakly Supervised Object Localization
=======
- pytorch 1.8.0
- opencv-contrib-python 4.5.2.54
- opencv-python 4.5.3.56
- Pillow 8.2.0
Use the test script to generate attention maps of LayerCAM, SGL, DGL on official pytorch VGG model
python test_loc.py conv5_1
Results on caffe VGG model using layercam_loc.
Rssults on different layers with 0.15 threshold.
| Layers | Method | Top1 loc | Top5 loc | GT-Known loc |
|---|---|---|---|---|
| conv5_3 | Layercam | 44.19 | 55.02 | 59.29 |
| conv5_3 | sgl-g1 | 32.16 | 39.90 | 42.80 |
| conv5_2 | Layercam | 45.83 | 56.88 | 61.23 |
| conv5_2 | sgl-g1 | 40.16 | 50.11 | 54.01 |
| conv5_1 | Layercam | 43.69 | 54.21 | 58.16 |
| conv5_1 | sgl-g1 | 47.70 | 59.32 | 63.87( 5.71)[best] |
| pool4 | Layercam | 44.10 | 54.83 | 58.99 |
| pool4 | sgl-g1 | 46.81 | 58.23 | 62.71 |
| conv4_3 | Layercam | 37.95 | 47.67 | 51.91 |
| conv4_3 | sgl-g1 | 53.91 | 43.45 | 57.82 |
| conv4_2 | Layercam | 42.85 | 53.54 | 57.89 |
| conv4_2 | sgl-g1 | 42.64 | 53.39 | 58.12 |
| conv4_1 | Layercam | 42.99 | 53.57 | 57.73 |
| conv4_1 | sgl-g1 | 39.08 | 49.09 | 53.48 |
Release codes about G2M