FreeAnchor Save
FreeAnchor: Learning to Match Anchors for Visual Object Detection (NeurIPS 2019)
FreeAnchor
The Code for "FreeAnchor: Learning to Match Anchors for Visual Object Detection".
This repository is based on maskrcnn-benchmark, and FreeAnchor has also been implemented in mmdetection, thanks @yhcao6 and @hellock.
New performance on COCO
We added multi-scale testing support and updated experiments. The previous version is in this branch.
| Backbone | Iteration | Training scales | Multi-scale testing |
AP (minival) |
AP (test-dev) |
Model |
|---|---|---|---|---|---|---|
| ResNet-50-FPN | 90k | 800 | N | 38.7 | 38.7 | Link |
| ResNet-101-FPN | 90k | 800 | N | 40.5 | 40.9 | Link |
| ResNet-101-FPN | 180k | [640, 800] | N | 42.7 | 43.1 | Link |
| ResNet-101-FPN | 180k | [480, 960] | N | 43.2 | 43.9 | Link |
| ResNet-101-FPN | 180k | [480, 960] | Y | 44.7 | 45.2 | Link |
| ResNeXt-64x4d-101-FPN | 180k | [640, 800] | N | 44.5 | 44.9 | Link |
| ResNeXt-64x4d-101-FPN | 180k | [480, 960] | N | 45.6 | 46.0 | Link |
| ResNeXt-64x4d-101-FPN | 180k | [480, 960] | Y | 46.8 | 47.3 | Link |
Notes:
- We use 8 GPUs with 2 image / GPU.
- In multi-scale testing, we use image scales in {480, 640, 800, 960, 1120, 1280} and max_size are 1.666× than scales.
Installation
Check INSTALL.md for installation instructions.
Usage
You will need to download the COCO dataset and configure your own paths to the datasets.
For that, all you need to do is to modify maskrcnn_benchmark/config/paths_catalog.py to point to the location where your dataset is stored.
Config Files
We provide four configuration files in the configs directory.
| Config File | Backbone | Iteration | Training scales |
|---|---|---|---|
| configs/free_anchor_R-50-FPN_1x.yaml | ResNet-50-FPN | 90k | 800 |
| configs/free_anchor_R-101-FPN_1x.yaml | ResNet-101-FPN | 90k | 800 |
| configs/free_anchor_R-101-FPN_j2x.yaml | ResNet-101-FPN | 180k | [640, 800] |
| configs/free_anchor_X-101-FPN_j2x.yaml | ResNeXt-64x4d-101-FPN | 180k | [640, 800] |
| configs/free_anchor_R-101-FPN_e2x.yaml | ResNet-101-FPN | 180k | [480, 960] |
| configs/free_anchor_X-101-FPN_e2x.yaml | ResNeXt-64x4d-101-FPN | 180k | [480, 960] |
Training with 8 GPUs
cd path_to_free_anchor
export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "path/to/config/file.yaml"
Test on COCO test-dev
cd path_to_free_anchor
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/test_net.py --config-file "path/to/config/file.yaml" MODEL.WEIGHT "path/to/.pth file" DATASETS.TEST "('coco_test-dev',)"
Multi-scale testing
cd path_to_free_anchor
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/multi_scale_test.py --config-file "path/to/config/file.yaml" MODEL.WEIGHT "path/to/.pth file" DATASETS.TEST "('coco_test-dev',)"
Evaluate NMS Recall
cd path_to_free_anchor
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/eval_NR.py --config-file "path/to/config/file.yaml" MODEL.WEIGHT "path/to/.pth file"
Citations
Please consider citing our paper in your publications if the project helps your research.
@inproceedings{zhang2019freeanchor,
title = {{FreeAnchor}: Learning to Match Anchors for Visual Object Detection},
author = {Zhang, Xiaosong and Wan, Fang and Liu, Chang and Ji, Rongrong and Ye, Qixiang},
booktitle = {Neural Information Processing Systems},
year = {2019}
}
Open Source Agenda Badge
