An unofficial implementation of RetinaNet in pytorch. Focal Loss for Dense Object Detection.
https://arxiv.org/abs/1708.02002
This repo achieves 35.1% mAP at nearly 700px resolution with a Resnet-50 backbone.
| paper (600px) | detectron (800px) | ours (nearly 700px) |
|---|---|---|
| 34 | 35.7 | 35.1 |
First, configure train.json file, add your root.
{
"root_train": "/home/xyt/dataset/VOC0712_trainval/JPEGImages",
"root_eval": "/home/xyt/dataset/VOCdevkit/VOC2007/JPEGImages",
"list_train": "data/voc_trainval.txt",
"list_eval": "data/voc_test.txt",
"name_file": "data/voc_name.txt",
"load": false,
"save": true,
"pretrain": true,
"freeze_bn": true,
"freeze_stages": 1,
"epoches": 30,
"nbatch_train": 28,
"nbatch_eval": 28,
"device": [0,1,2,3],
"num_workers": 14,
"lr_base": 0.01,
"lr_gamma": 0.1,
"lr_schedule": [16000, 22000],
"momentum": 0.9,
"weight_decay": 0.0001,
"boxarea_th": 32,
"grad_clip": 3,
"img_scale_min": 0.6,
"augmentation": false
}Then, configure some parameters in detector.py file.
# TODO: choose backbone
from backbone import resnet50 as backbone
# TODO: configure Detector
self.view_size = 641
self.classes = 20 # TODO: total 20 classes exclude backgroundIn my experiment, only 30 epochs were performed. Better results can be achieved if it takes longer. run train to get results. It takes about 5 hours with 4x Titan-Xp. run analyze and got mAP@.5: 79.1%
map_mean
[0.0372 0.1881 0.312 0.3404 0.3744 0.3819 0.3969 0.4198 0.4351 0.452
0.4424 0.4602 0.4708 0.4692 0.4735 0.4725 0.4683 0.4809 0.4755 0.4909
0.5218 0.5249 0.525 0.5219 0.5261 0.5251 0.524 0.5249 0.5247 0.5258]
map_50
[0.0861 0.3869 0.5774 0.6091 0.6529 0.6574 0.6697 0.6924 0.7126 0.7339
0.7158 0.7387 0.7425 0.7427 0.7431 0.7493 0.7435 0.7473 0.7394 0.7567
0.7884 0.7925 0.7889 0.7878 0.792 0.792 0.7868 0.7899 0.7905 0.7909]
map_75
[0.026 0.1617 0.2952 0.3366 0.3847 0.393 0.4169 0.4451 0.4621 0.4809
0.47 0.4879 0.5032 0.5026 0.5053 0.5108 0.502 0.5129 0.5166 0.5329
0.5632 0.5661 0.5642 0.5654 0.5667 0.5661 0.5654 0.565 0.5654 0.5656]First, configure train.json file, add your root.
{
"root_train": "/home1/xyt/dataset/coco17/images",
"root_eval": "/home1/xyt/dataset/coco17/images",
"list_train": "data/coco_train2017.txt",
"list_eval": "data/coco_val2017.txt",
"name_file": "data/coco_name.txt",
"load": false,
"save": true,
"pretrain": true,
"freeze_bn": true,
"freeze_stages": 1,
"epoches": 12,
"nbatch_train": 16,
"nbatch_eval": 16,
"device": [1,2,3,5,6,7,8,9],
"num_workers": 16,
"lr_base": 0.01,
"lr_gamma": 0.1,
"lr_schedule": [60000, 80000],
"momentum": 0.9,
"weight_decay": 0.0001,
"boxarea_th": 32,
"grad_clip": 3,
"img_scale_min": 0.8,
"augmentation": false
}Then, configure some parameters in detector.py file.
self.view_size = 1025
self.classes = 80 # TODO: total 80 classes exclude backgroundIt takes about 15 hours with 8x Titan-Xp. Run analyze to get mAP curves.
map_mean
[0.1401 0.2009 0.2189 0.2455 0.2534 0.2556 0.2672 0.2726 0.3224 0.3291
0.3329 0.3343]
map_50
[0.2548 0.3444 0.3679 0.4026 0.4157 0.4158 0.4316 0.4363 0.5043 0.5142
0.5174 0.5193]
map_75
[0.1392 0.2068 0.2273 0.257 0.2668 0.2696 0.2831 0.2906 0.3442 0.3549
0.3576 0.3588]Run cocoeval and got mAP: 33.7%
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.337
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.524
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.361
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.178
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.370
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.445
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.277
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.432
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.462
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.275
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.501
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.583Like 2 in train.json modify key
"epoches": 24,
"lr_schedule": [120000, 160000],Run train to get results. It takes about 30 hours with 8x Titan-Xp. Run analyze to get mAP curves.
map_mean
[0.1373 0.2022 0.2218 0.2404 0.2539 0.2638 0.26 0.2638 0.2783 0.2777
0.2791 0.2865 0.285 0.2762 0.2888 0.2982 0.3398 0.3424 0.345 0.3474
0.3459 0.3489 0.3494 0.3489]
map_50
[0.2521 0.3387 0.3788 0.3994 0.4142 0.426 0.4256 0.4268 0.446 0.4431
0.4511 0.462 0.4533 0.4433 0.4621 0.4688 0.5253 0.5299 0.5315 0.5354
0.5335 0.5375 0.538 0.5378]
map_75
[0.1364 0.213 0.2323 0.2508 0.2689 0.2782 0.2765 0.2801 0.2952 0.2973
0.2955 0.3054 0.2996 0.2947 0.3091 0.3196 0.3636 0.3666 0.3691 0.3726
0.37 0.3725 0.3739 0.3731]Run cocoeval and got mAP: 35.1%
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.351
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.541
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.374
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.189
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.385
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.464
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.288
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.445
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.474
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.291
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.519
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.598new-1x
map_mean
[0.1471 0.2116 0.2318 0.2544]
map_50
[0.271 0.3581 0.3845 0.4155]
map_75
[0.1448 0.2176 0.2432 0.2707]