LSTFE-Net:Long Short-Term Feature Enhancement Network for Video Small Object Detection

May 21, 2024 ยท View on GitHub

This repo is an official implementation of "LSTFE-net: Long short-term feature enhancement network for video small object detection", accepted by CVPR 2023.

Citing LSTFE

Please cite our paper in your publications if it helps your research:

@inproceedings{xiao2023lstfe,
  title={LSTFE-net: Long short-term feature enhancement network for video small object detection},
  author={Xiao, Jinsheng and Wu, Yuanxu and Chen, Yunhua and Wang, Shurui and Wang, Zhongyuan and Ma, Jiayi},
  booktitle={the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={14613--14622},
  year={2023}
}

Installation

Please follow INSTALL.md for installation instructions.

Data preparation

Please download FLDrones dataset. After that, we recommend to symlink the path to the datasets to datasets/. And the path structure should be as follows:

./datasets/FLDrones/
./datasets/FLDrones/Annotations/VID
./datasets/FLDrones/Data/VID
./datasets/FLDrones/ImageSets

Inference

The inference command line for testing on the validation dataset:

python -m torch.distributed.run \
    --nproc_per_node 1 \
    tools/test_net.py \
    --config-file configs/LSTFE/vid_R_101_C4_LSTFE_1x.yaml \
    MODEL.WEIGHT FLDrones_lstfe.pth

Please note that:

  1. FLdrones_lstfe.pth is your model name
  2. If you want to evaluate a different model, please change --config-file to its config file and MODEL.WEIGHT to its weights file.

Training

The following command line will train LSTFE_Resnet101 on 1 GPUs with Synchronous Stochastic Gradient Descent (SGD):

python -m torch.distributed.run \
    --nproc_per_node 1 \
    tools/train_net.py \
    --config-file configs/LSTFE/vid_R_101_C4_LSTFE_1x.yaml \
    OUTPUT_DIR training_dir/LSTFE

Please note that:

  1. The models will be saved into OUTPUT_DIR.

Contributing to the project

Any pull requests or issues are welcomed.