Development, Maintenance @ChenJoya. Please feel free to contact me: [email protected]

To address the foreground-background imbalance, is sampling heuristics necessary in training deep object detectors?

Keep clam and try the sampling-free mechanism in this repository.

Sampling-free mechanism enables various object detectors (e.g. one-stage, two-stage, anchor-free, multi-stage) to drop sampling heuristics (e.g., undersampling, Focal Loss, objectness), but achieve better bounding-box or instance segmentation accuracy.

Technical report: https://arxiv.org/abs/1909.04868. This repository is based on maskrcnn-benchmark, including the implementation of RetinaNet, Faster/Mask R-CNN. Other detectors will also be released.

Check INSTALL.md for installation instructions.

Check datasets/README.md for preparing your datasets.

By setting sampling_free_on=True in scripts/train.sh, you can easily train with the sampling-free mechanism.

We give an example of training Faster R-CNN with the sampling-free mechanism on COCO, using 4 GPUs.

1. Modifying the scripts/train.sh：

det=faster_rcnn
backbone=R_50_FPN
schedule=0.2x

sampling_free_on=True
# 1.0 for retinanet, 2.0 for rcnn
classification_scale=2.0

gpus=0,1,2,3
gpun=4
master_addr=127.0.0.1
master_port=29501

2. Executing scripts/train.sh：

./scripts/train.sh

3. Now the training is done. Your work dir is in backup/your_model_name/, with every 5000 iterations saving a model.

By modifying scripts/eval.sh, you can easily evaluate your trained model.

We give an example of evaluating trained model using 4 GPUs, with 16 images per GPU.

1. Modifying the scripts/eval.sh:

model=faster_rcnn_voc_R_50_FPN_0.2x_freex2.0
network=faster_rcnn_voc_R_50_FPN

gpus=0,1,2,3
gpun=4
master_addr=127.0.0.8
master_port=29508

# suitable for 1080Ti (12G)
# change its value for your GPU
ims_per_batch=$[16*$gpun]

Here, network=faster_rcnn_voc_R_50_FPN refers to the model name without schedule.

2. Setting the iterations that you want to evaluate in scripts/eval.sh:

for iter in "0060000" "0090000"：
  ......

3. Executing scripts/eval.sh：

./scripts/eval.sh

Pre-trained models, trained baselines can be found in MODEL_ZOO.md.

See the original benchmark maskrcnn-benchmark for more details.

Please consider citing this project in your publications if it helps your research. The following is a BibTeX reference. The BibTeX entry requires the url LaTeX package.

@article{sampling_free,
author    = {Joya Chen and
             Dong Liu and
             Tong Xu and
             Shilong Zhang and
             Shiwei Wu and
             Bin Luo and
             Xuezheng Peng and
             Enhong Chen},
title     = {Is Sampling Heuristics Necessary in Training Deep Object Detectors?},
journal   = {CoRR},
volume    = {abs/1909.04868},
year      = {2019},
url       = {http://arxiv.org/abs/1909.04868},
archivePrefix = {arXiv},
eprint    = {1909.04868},
}

sampling-free is released under the MIT license. See LICENSE for additional details.