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;
Zhi Tian, 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.
Implementation based on Detectron2 is included in AdelaiDet.
A real-time model with 46FPS and 40.3 in AP on COCO minival is also available here.
- Totally anchor-free: FCOS completely avoids the complicated computation related to anchor boxes and all hyper-parameters of anchor boxes.
- Better performance: The very simple one-stage detector achieves much better performance (38.7 vs. 36.8 in AP with ResNet-50) than Faster R-CNN. Check out more models and experimental results here.
- Faster training and testing: With the same hardwares and backbone ResNet-50-FPN, FCOS also requires less training hours (6.5h vs. 8.8h) than Faster R-CNN. FCOS also takes 12ms less inference time per image than Faster R-CNN (44ms vs. 56ms).
- State-of-the-art performance: Our best model based on ResNeXt-64x4d-101 and deformable convolutions achieves 49.0% in AP on COCO test-dev (with multi-scale testing).
- FCOS with Fast And Diverse (FAD) neural architecture search is avaliable at FAD. (30/10/2020)
- Script for exporting ONNX models. (21/11/2019)
- New NMS (see #165) speeds up ResNe(x)t based models by up to 30% and MobileNet based models by 40%, with exactly the same performance. Check out here. (12/10/2019)
- New models with much improved performance are released. The best model achieves 49% in AP on COCO test-dev with multi-scale testing. (11/09/2019)
- FCOS with VoVNet backbones is available at VoVNet-FCOS. (08/08/2019)
- A trick of using a small central region of the BBox for training improves AP by nearly 1 point as shown here. (23/07/2019)
- FCOS with HRNet backbones is available at HRNet-FCOS. (03/07/2019)
- FCOS with AutoML searched FPN (R50, R101, ResNeXt101 and MobileNetV2 backbones) is available at NAS-FCOS. (30/06/2019)
- FCOS has been implemented in mmdetection. Many thanks to @yhcao6 and @hellock. (17/05/2019)
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 git+https://github.com/tianzhi0549/FCOS.git
# 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://huggingface.co/tianzhi/FCOS/resolve/main/FCOS_imprv_R_50_FPN_1x.pth?download=true -O FCOS_imprv_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_imprv_R_50_FPN_1x.yaml \
MODEL.WEIGHT FCOS_imprv_R_50_FPN_1x.pth \
TEST.IMS_PER_BATCH 4
Please note that:
- If your model's name is different, please replace
FCOS_imprv_R_50_FPN_1x.pth
with your own. - If you enounter out-of-memory error, please try to reduce
TEST.IMS_PER_BATCH
to 1. - If you want to evaluate a different model, please change
--config-file
to its config file (in configs/fcos) andMODEL.WEIGHT
to its weights file. - Multi-GPU inference is available, please refer to #78.
- We improved the postprocess efficiency by using multi-label nms (see #165), which saves 18ms on average. The inference metric in the following tables has been updated accordingly.
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 | Multi-scale training | Testing time / im | AP (minival) | Link |
---|---|---|---|---|
FCOS_imprv_R_50_FPN_1x | No | 44ms | 38.7 | download |
FCOS_imprv_dcnv2_R_50_FPN_1x | No | 54ms | 42.3 | download |
FCOS_imprv_R_101_FPN_2x | Yes | 57ms | 43.0 | download |
FCOS_imprv_dcnv2_R_101_FPN_2x | Yes | 73ms | 45.6 | download |
FCOS_imprv_X_101_32x8d_FPN_2x | Yes | 110ms | 44.0 | download |
FCOS_imprv_dcnv2_X_101_32x8d_FPN_2x | Yes | 143ms | 46.4 | download |
FCOS_imprv_X_101_64x4d_FPN_2x | Yes | 112ms | 44.7 | download |
FCOS_imprv_dcnv2_X_101_64x4d_FPN_2x | Yes | 144ms | 46.6 | download |
Note that imprv
denotes improvements
in our paper Table 3. These almost cost-free changes improve the performance by ~1.5% in total. Thus, we highly recommend to use them. The following are the original models presented in our initial paper.
Model | Multi-scale training | Testing time / im | AP (minival) | AP (test-dev) | Link |
---|---|---|---|---|---|
FCOS_R_50_FPN_1x | No | 45ms | 37.1 | 37.4 | download |
FCOS_R_101_FPN_2x | Yes | 59ms | 41.4 | 41.5 | download |
FCOS_X_101_32x8d_FPN_2x | Yes | 110ms | 42.5 | 42.7 | download |
FCOS_X_101_64x4d_FPN_2x | Yes | 113ms | 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 | 26ms | 30.9 | download |
FCOS_syncbn_bs32_MNV2_FPN_1x | 32 | No | 33ms | 33.1 | download |
FCOS_bn_bs16_MNV2_FPN_1x | 16 | No | 44ms | 31.0 | download |
[1] 1x and 2x mean the model is trained for 90K and 180K iterations, respectively.
[2] All results are obtained with a single model and without any test time data augmentation such as multi-scale, flipping and etc..
[3] c128
denotes the model has 128 (instead of 256) channels in towers (i.e., MODEL.RESNETS.BACKBONE_OUT_CHANNELS
in config).
[4] dcnv2
denotes deformable convolutional networks v2. Note that for ResNet based models, we apply deformable convolutions from stage c3 to c5 in backbones. For ResNeXt based models, only stage c4 and c5 use deformable convolutions. All models use deformable convolutions in the last layer of detector towers.
[5] The model FCOS_imprv_dcnv2_X_101_64x4d_FPN_2x
with multi-scale testing achieves 49.0% in AP on COCO test-dev. Please use TEST.BBOX_AUG.ENABLED True
to enable multi-scale testing.
The following command line will train FCOS_imprv_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 \
--config-file configs/fcos/fcos_imprv_R_50_FPN_1x.yaml \
DATALOADER.NUM_WORKERS 2 \
OUTPUT_DIR training_dir/fcos_imprv_R_50_FPN_1x
Note that:
- If you want to use fewer GPUs, please change
--nproc_per_node
to 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_BATCH
in 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
. - If you want to train FCOS on your own dataset, please follow this instruction #54.
- Now, training with 8 GPUs and 4 GPUs can have the same performance. Previous performance gap was because we did not synchronize
num_pos
between GPUs when computing loss.
Please refer to the directory onnx for an example of exporting the model to ONNX. A converted model can be downloaded here. We recommend you to use PyTorch >= 1.4.0 (or nightly) and torchvision >= 0.5.0 (or nightly) for ONNX models.
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}
}
@article{tian2021fcos,
title = {{FCOS}: A Simple and Strong Anchor-free Object Detector},
author = {Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
booktitle = {IEEE T. Pattern Analysis and Machine Intelligence (TPAMI)},
year = {2021}
}
We would like to thank @yqyao for the tricks of center sampling and GIoU. We also thank @bearcatt for his suggestion of positioning the center-ness branch with box regression (refer to #89).
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.