xytpai / fcos

An unofficial implementation of FCOS in Pytorch: 37.3-resnet50-700px

FCOS in Pytorch

An unofficial implementation of FCOS in Pytorch. FCOS: Fully Convolutional One-Stage Object Detection.

https://arxiv.org/abs/1904.01355

Some modifies are adopted:

• Remove center-ness branch for simplicity.
• Add center sample mechanism to improve performance.
• Predect yxhw instead of tlbr +0.3Map.
• Note: GPU Compute Capability >= 6.1 (pytorch>=1.0.0)

Download:

• Baidu https://pan.baidu.com/s/1MjsgE2fu86qmWZpVD7a8Aw
• Google https://drive.google.com/open?id=1U-jwZV54ydiuLR5zcOzdbGDXUNDKJZtU

paper (800px)	official (800px)	ours (nearly 700px)
36.6	38.7	37.3

Some cuda operations are provided for acceleration.

cd libs/nms_cuda
python setup.py install
cd libs/sigmoid_focal_loss_cuda
python setup.py install
cd libs/assign_box_cuda
python setup.py install

Inference

1. Copy your images (xxx.jpg) to /images.
2. Command: python run_demo.py

COCO (2x)

1. Copy configs/detector_coco.py to ./detector.py
2. Copy configs/train_coco_2x.json to ./train.json
3. Configure train.json file, add your root.
4. Command: python run_train.py to start schedule, it takes about 30 hours with 8x Titan-XP.
5. Command: python run_analyze.py to get mAP curves.
6. Command: python run_cocoeval.py to get mAP.

map_mean
[0.1345 0.1984 0.2192 0.2468 0.2496 0.2606 0.2729 0.2708 0.2768 0.2872
 0.2924 0.295  0.2973 0.3041 0.2927 0.3027 0.3565 0.3583 0.3622 0.3642
 0.3659 0.3679 0.3689 0.37  ]
map_50
[0.2482 0.3404 0.3749 0.3997 0.4059 0.4182 0.4401 0.4393 0.4389 0.4496
 0.4584 0.4646 0.4618 0.4719 0.4584 0.4741 0.535  0.5385 0.5405 0.5436
 0.5457 0.5489 0.5497 0.5511]
map_75
[0.1304 0.2044 0.2265 0.2596 0.2595 0.2758 0.2895 0.283  0.2914 0.3037
 0.311  0.315  0.3145 0.3246 0.3136 0.3232 0.3804 0.3829 0.3869 0.3906
 0.3917 0.3924 0.3943 0.3962]

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.373
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.556
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.398
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.198
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.411
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.486
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.303
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.482
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.525
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.326
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.575
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.639

VOC0712 (30 epoches)

1. Copy configs/detector_voc.py to ./detector.py
2. Copy configs/train_voc.json to ./train.json
3. Configure train.json file, add your root.
4. Command: python run_train.py to start schedule.
5. Command: python run_analyze.py to get mAP curves.

map_mean
[0.0315 0.174  0.2992 0.3442 0.3567 0.3983 0.4329 0.4475 0.44   0.4525
 0.4397 0.472  0.4743 0.475  0.464  0.4885 0.4855 0.4908 0.4923 0.4776
 0.529  0.5331 0.535  0.5375 0.5371 0.5348 0.538  0.537  0.5378 0.5368]
map_50
[0.0721 0.3466 0.5524 0.594  0.6051 0.6488 0.6968 0.7062 0.7001 0.7205
 0.6982 0.7211 0.7258 0.7276 0.7092 0.7408 0.7381 0.7362 0.7295 0.7246
 0.7682 0.7718 0.7736 0.7771 0.7741 0.7725 0.7767 0.7737 0.7744 0.7732]
map_75
[0.0219 0.1561 0.2932 0.3555 0.3639 0.4224 0.4618 0.4781 0.4718 0.4849
 0.479  0.5084 0.5147 0.5116 0.4989 0.5293 0.5259 0.5244 0.5292 0.5174
 0.574  0.5789 0.5797 0.5792 0.5794 0.5805 0.5858 0.5814 0.5829 0.5819]

How to change resolution ?

To change the prediction resolution, you need to modify the following parameters. Note that the resolution must be 128*n+1 (n=1,2,3...).

# ---------------------------
self.view_size = 1025 # 128*n+1
self.phpw = [[129, 129], [65, 65], [33, 33], [17, 17], [9, 9]]
# [[16n+1, 16n+1], [8n+1, 8n+1], [4n+1, 4n+1], [2n+1, 2n+1], [n+1, n+1]]
# ---------------------------

How to train your own data ?

1. Create data_name.txt

background
classname1
classname2
classname3

2. Create data_train.txt

xxx1.jpg 237 411 480 504 1 313 9 391 151 3
xxx2.jpg 261 234 553 134 2
# img_name ymin xmin ymax xmax class_id ymin xmin ymax xmax class_id ... \n 

3. Create data_eval.txt (same format as data_train.txt)
4. Config train.json file, add your root.

Call for suggestions !

We can't achieve the above results in standard 12-epoch training, even if we add the center branch. The possible reasons are: 1. Instead of set minimum side to 800, we set the maxmal side to 1025. 2. There are some tricks we haven't adopted.

fcos's Issues

在pascal上的准确率

我该怎么修改代码得到pascal上的准确率呢

Train on the VOC dataset

If using VOC to train ,how can I set the learning rate(2 GPUs)? Looking forward for your reply,thank you very much

警告问题 warning problem

indexing with dtype torch.uint8 is now deprecated, please use a dtype torch.bool instead.
how to avoid this warning when training？

about accuracy in coco?

映射回原图操作的问题

你好，请问映射回原图的操作是在代码哪里体现的？我看到之前有人提问提到过c_y, c_x = h * float(stride), w * float(stride)，请问在代码哪个部分？

关于centerness

您好，请问你把centerness去掉后，不会引起误检测变多么？我看原文中这是一个创新点

assign box中的参数

你好，请问如果修改训练时候图像的输入尺寸，1024，200,5这几个数字需要修改吗？比如要修改成256*896大小训练

iou_loss的计算问题

回归的输出不是偏移量吗，iou_loss的计算是否有问题

About the assign_box()

Hello author, I found that in the assign_box.py, the code without the assign_box_cuda is not given, can I ask for the function assign_box(), and the code that can be used without cuda.
Thanks a lot!

about training setting

I find that the official code get 38.7%AP(80k, about 12 epoch)
I modify your train.json by following official training instructions(12 epoch) and then only get 30%AP

训练尺寸问题

代码默认的尺寸是1025*1025，而不是readme里写的700？

Segmentation fault (core dumped)

运行run_train.py后，提示

use lib assign_box_cuda
use lib sigmoid_focal_loss_cuda
use lib nms_cuda
Segmentation fault (core dumped)

调试后定位到net = net.cuda(cfg['device'][0])
这是CUDA的版本问题吗？

Exception: not find lib assign_box_cuda

fcos for geospatial

Hello, I'm trying to your FCOS implementaitn on my own geospatial dataset (I guess it's similar to the spacenet dataset) to detect buildings. I've followed your instructions for setting up the training and evaluation datasets, but I get no output, (the loss seems to go down but i'm guessing it's converging on predicting nothing).

Just to confirm a few silly things. Your xmin, xmax, ymin, ymax are defined in the standard format right? So left, right, top, bottom.

The typical image i'm training on looks like this:

Do you have any ideas of what parameters I might need to adjust to get good output?

Thanks a lot!
Roger

关于精度和训练问题

readme 里的37的精度是 2x 的训练结果？还有4 GPU 的精度是多少，我看跑了官方的源码resnet50 4GPU训练1x 有38+的精度

Alternate download link for the pretrained net.pkl

Hello,
could you please upload the net.pkl model file hosted on baidu (https://pan.baidu.com/s/1MjsgE2fu86qmWZpVD7a8Aw) to another file hoster, or maybe as a github artifact? It is pretty cumbersome to download from baidu without registration or additional client software.

Thanks a lot.

请问你有没有在voc把输入设为416x416下训练过？

映射回原图的坐标问题

请问c_y, c_x = h * float(stride), w * float(stride)在映射回原图的方法为什么和fcos论文提到的ws+1/2s和hs+1/2s不一样呢？

训练coco时候，focal loss不能正常收敛

你好，我用这个代码，训练coco2017数据集，发现focal loss不能正常收敛，完全偏向了负样本，正样本的分类准确率为0。我使用python3.7和pytorch1.0，batch size为8。请问你有什么经验和建议么

loss:Nan

Loss printed out is NAN

关于参数的一些问题

关于run_train.py中net指定gpu的问题

net = torch.nn.DataParallel(net, device_ids=cfg['device'])
net = net.cuda(cfg['device'][0])
net.train()
这是在run_train.py中准备model时候的操作。
想请问，为什么先分配到所有的gpu上net = torch.nn.DataParallel(net, device_ids=cfg['device'])，然后再单独指定一块net = net.cuda(cfg['device'][0])？不太清楚其中意义。

希望能够获得答复，谢谢！