该代码与FCOS原代码一样，我为了方便自己理解，给整个训练过程和测试过程添加了部分注释．已经包括数据集的读取和真值读取部分． 希望可以帮到大家．一同来进行理解
This project hosts the code for implementing the FCOS algorithm for object detection, as presented in our paper:
FCOS: Fully Convolutional One-Stage Object Detection;
Tian Zhi, Chunhua Shen, Hao Chen, and Tong He;
In: Proc. Int. Conf. Computer Vision (ICCV), 2019.
arXiv preprint arXiv:1904.01355
The full paper is available at: https://arxiv.org/abs/1904.01355.
- Totally anchor-free: FCOS completely avoids the complicated computation related to anchor boxes and all hyper-parameters of anchor boxes.
- Memory-efficient: FCOS uses 2x less training memory footprint than its anchor-based counterpart RetinaNet.
- Better performance: The very simple detector achieves better performance (37.1 vs. 36.8) than Faster R-CNN.
- Faster training: With the same hardwares, FCOS also requires less training hours (6.5h vs. 8.8h) than Faster R-CNN.
- State-of-the-art performance: Without bells and whistles, FCOS achieves state-of-the-art performances. It achieves 41.5% (ResNet-101-FPN) and 43.2% (ResNeXt-64x4d-101) in AP on coco test-dev.
- FCOS with VoVNet backbones is available at VoVNet-FCOS.
- A trick of using a small central region of the BBox for training improves AP by nearly 1 point as shown here.
- FCOS with HRNet backbones is available at HRNet-FCOS.
- FCOS with AutoML searched FPN (R50, R101, ResNeXt101 and MobileNetV2 backbones) is available at NAS-FCOS.
- FCOS has been implemented in mmdetection. Many thanks to @yhcao6 and @hellock.
We use 8 Nvidia V100 GPUs.
But 4 1080Ti GPUs can also train a fully-fledged ResNet-50-FPN based FCOS since FCOS is memory-efficient.
For users who only want to use FCOS as an object detector in their projects, they can install it by pip. To do so, run:
pip install torch # install pytorch if you do not have it
pip install fcos
# run this command line for a demo
fcos https://github.com/tianzhi0549/FCOS/raw/master/demo/images/COCO_val2014_000000000885.jpg
Please check out here for the interface usage.
This FCOS implementation is based on maskrcnn-benchmark. Therefore the installation is the same as original maskrcnn-benchmark.
Please check INSTALL.md for installation instructions. You may also want to see the original README.md of maskrcnn-benchmark.
Once the installation is done, you can follow the below steps to run a quick demo.
# assume that you are under the root directory of this project,
# and you have activated your virtual environment if needed.
wget https://cloudstor.aarnet.edu.au/plus/s/dDeDPBLEAt19Xrl/download -O FCOS_R_50_FPN_1x.pth
python demo/fcos_demo.py
The inference command line on coco minival split:
python tools/test_net.py \
--config-file configs/fcos/fcos_R_50_FPN_1x.yaml \
MODEL.WEIGHT FCOS_R_50_FPN_1x.pth \
TEST.IMS_PER_BATCH 4
Please note that:
- If your model's name is different, please replace
FCOS_R_50_FPN_1x.pthwith your own. - If you enounter out-of-memory error, please try to reduce
TEST.IMS_PER_BATCHto 1. - If you want to evaluate a different model, please change
--config-fileto its config file (in configs/fcos) andMODEL.WEIGHTto its weights file.
For your convenience, we provide the following trained models (more models are coming soon).
ResNe(x)ts:
All ResNe(x)t based models are trained with 16 images in a mini-batch and frozen batch normalization (i.e., consistent with models in maskrcnn_benchmark).
| Model | Total training mem (GB) | Multi-scale training | Testing time / im | AP (minival) | AP (test-dev) | Link |
|---|---|---|---|---|---|---|
| FCOS_R_50_FPN_1x | 29.3 | No | 71ms | 37.1 | 37.4 | download |
| FCOS_R_101_FPN_2x | 44.1 | Yes | 74ms | 41.4 | 41.5 | download |
| FCOS_X_101_32x8d_FPN_2x | 72.9 | Yes | 122ms | 42.5 | 42.7 | download |
| FCOS_X_101_64x4d_FPN_2x | 77.7 | Yes | 140ms | 43.0 | 43.2 | download |
MobileNets:
We update batch normalization for MobileNet based models. If you want to use SyncBN, please install pytorch 1.1 or later.
| Model | Training batch size | Multi-scale training | Testing time / im | AP (minival) | Link |
|---|---|---|---|---|---|
| FCOS_syncbn_bs32_c128_MNV2_FPN_1x | 32 | No | 19ms | 30.9 | download |
| FCOS_syncbn_bs32_MNV2_FPN_1x | 32 | No | 59ms | 33.1 | download |
| FCOS_bn_bs16_MNV2_FPN_1x | 16 | No | 59ms | 31.0 | download |
[1] 1x and 2x mean the model is trained for 90K and 180K iterations, respectively.
[2] We report total training memory footprint on all GPUs instead of the memory footprint per GPU as in maskrcnn-benchmark.
[3] All results are obtained with a single model and without any test time data augmentation such as multi-scale, flipping and etc..
[4] Our results have been improved since our initial release. If you want to check out our original results, please checkout commit f4fd589.
[5] c128 denotes the model has 128 (instead of 256) channels in towers (i.e., MODEL.RESNETS.BACKBONE_OUT_CHANNELS in config).
The following command line will train FCOS_R_50_FPN_1x on 8 GPUs with Synchronous Stochastic Gradient Descent (SGD):
python -m torch.distributed.launch \
--nproc_per_node=8 \
--master_port=$((RANDOM + 10000)) \
tools/train_net.py \
--skip-test \
--config-file configs/fcos/fcos_R_50_FPN_1x.yaml \
DATALOADER.NUM_WORKERS 2 \
OUTPUT_DIR training_dir/fcos_R_50_FPN_1x
Note that:
- If you want to use fewer GPUs, please change
--nproc_per_nodeto the number of GPUs. No other settings need to be changed. The total batch size does not depends onnproc_per_node. If you want to change the total batch size, please changeSOLVER.IMS_PER_BATCHin configs/fcos/fcos_R_50_FPN_1x.yaml. - The models will be saved into
OUTPUT_DIR. - If you want to train FCOS with other backbones, please change
--config-file. - The link of ImageNet pre-training X-101-64x4d in the code is invalid. Please download the model here.
- If you want to train FCOS on your own dataset, please follow this instruction #54.
Any pull requests or issues are welcome.
Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.
@inproceedings{tian2019fcos,
title = {{FCOS}: Fully Convolutional One-Stage Object Detection},
author = {Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
booktitle = {Proc. Int. Conf. Computer Vision (ICCV)},
year = {2019}
}
For academic use, this project is licensed under the 2-clause BSD License - see the LICENSE file for details. For commercial use, please contact the authors.