Avg 0.174921s per iteration
Overlap_threshold=0.3, overall(12032 objs), MAP=0.670017
Overlap_threshold=0.4, overall(12032 objs), MAP=0.625412
Overlap_threshold=0.5, overall(12032 objs), MAP=0.544122
Avg 0.147702s per iteration
Overlap_threshold=0.3, overall(12032 objs), MAP=0.398297
Overlap_threshold=0.4, overall(12032 objs), MAP=0.342526
Overlap_threshold=0.5, overall(12032 objs), MAP=0.257104
| Dataset 1: 822 categories, from VIG Brand1500, Flickr32 Logo, Flickr27 Logo, Belga Logo dataset, | |
| Processing strategy: | |
| 1. keep images containing at least one medium/large size logo (>=32*32 pixels) | |
| 2. remove categories with containing less than 40 logos after step 1. | |
| Results on overall objects (11044) | |
| | IOU | MAP | Th@0.8 | Prec@0.8 | Rec@0.8 | Th@0.9 | Prec@0.9 | Rec@0.9 | Th@0.95 | Prec@0.95 | Rec@0.95 | | |
| | :-- | :----: | :----: | :------: | :-----: | :----: | :------: | :-----: | :-----: | :-------: | :------: | | |
| | 0.3 | 0.6249 | 0.7626 | 0.8 | 0.5872 | 0.9321 | 0.9001 | 0.4772 | 0.9891 | 0.9502 | 0.3022 | | |
| | 0.4 | 0.6073 | 0.7864 | 0.8 | 0.5666 | 0.9551 | 0.9001 | 0.4341 | 0.9963 | 0.9503 | 0.1868 | | |
| | 0.5 | 0.5721 | 0.8386 | 0.8 | 0.5235 | 0.9832 | 0.9001 | 0.3305 | 0.9995 | 0.9507 | 0.0367 | |
| Dataset 1: 1048 categories, | |
| 1. keep images containing at least one medium/large size logo (>=32*32 pixels) | |
| 2. Remove images which aspect ratio >4 or <0.25 | |
| 3. remove categories with containing less than 20 logos after step 2. | |
| 4. random rotate image from step 3 [-10,10] degree 3 times | |
| 5. merge the data from step 3 & 4 | |
| Results on overall objects (11545) | |
| | IOU | MAP | Th@0.8 | Prec@0.8 | Rec@0.8 | Th@0.9 | Prec@0.9 | Rec@0.9 | Th@0.95 | Prec@0.95 | Rec@0.95 | | |
| | :-- | :----: | :----: | :------: | :-----: | :----: | :------: | :-----: | :-----: | :-------: | :------: | |
| ## ZF | |
| | IOU | MAP | Th@0.8 | Prec@0.8 | Rec@0.8 | Th@0.9 | Prec@0.9 | Rec@0.9 | Th@0.95 | Prec@0.95 | Rec@0.95 | | |
| | :-- | :----: | :----: | :------: | :-----: | :----: | :------: | :-----: | :-----: | :-------: | :------: | | |
| | 0.3 | 0.7018 | 0.7302 | 0.8001 | 0.6708 | 0.9111 | 0.9001 | 0.5757 | 0.9704 | 0.9501 | 0.4715 | | |
| | 0.4 | 0.6693 | 0.7833 | 0.8001 | 0.6362 | 0.9359 | 0.9001 | 0.5345 | 0.981 | 0.9501 | 0.4205 | | |
| | 0.5 | 0.6114 | 0.8611 | 0.8001 | 0.5695 | 0.9721 | 0.9 | 0.44 | 0.9934 | 0.9502 | 0.3044 | | |
| ## Resnet 18bn | |
| | IOU | MAP | Th@0.8 | Prec@0.8 | Rec@0.8 | Th@0.9 | Prec@0.9 | Rec@0.9 | Th@0.95 | Prec@0.95 | Rec@0.95 | | |
| | :-- | :----: | :----: | :------: | :-----: | :----: | :------: | :-----: | :-----: | :-------: | :------: | |
| ZF model train.py -g 0,1,2,3 -d imagenet -a -n zf -t 28e -e 28e | |
| Results on small objects (5985) | |
| | IOU | MAP | Th@0.8 | Prec@0.8 | Rec@0.8 | Th@0.9 | Prec@0.9 | Rec@0.9 | Th@0.95 | Prec@0.95 | Rec@0.95 | | |
| | :-- | :----: | :----: | :------: | :-----: | :----: | :------: | :-----: | :-----: | :-------: | :------: | | |
| | 0.3 | 0.0073 | 1.0 | 1.0 | 0.0 | 1.0 | 1.0 | 0.0 | 1.0 | 1.0 | 0.0 | | |
| | 0.4 | 0.0063 | 1.0 | 1.0 | 0.0 | 1.0 | 1.0 | 0.0 | 1.0 | 1.0 | 0.0 | | |
| | 0.5 | 0.0049 | 1.0 | 1.0 | 0.0 | 1.0 | 1.0 | 0.0 | 1.0 | 1.0 | 0.0 | | |
| Results on large objects (33193) | |
| | IOU | MAP | Th@0.8 | Prec@0.8 | Rec@0.8 | Th@0.9 | Prec@0.9 | Rec@0.9 | Th@0.95 | Prec@0.95 | Rec@0.95 | | |
| | :-- | :----: | :----: | :------: | :-----: | :----: | :------: | :-----: | :-----: | :-------: | :------: | |
| ##Dense Scale Pyramid Resnet18 | |
| Baseline: 450000 iterations, batchsize 256 | |
| Resnet18 112 | |
| accuracy = 0.62748 | |
| accuracy_top_5 = 0.84526 | |
| Resnet18 224 | |
| accuracy = 0.6934 | |
| accuracy_top_5 = 0.8878 |