MVA2023 - Small Object Detection Challenge for Spotting Birds

Author: Kazuki Takasan (k-takasan)

This repository includes my submission code used in the challenge .

Overview of my solution

I performed an ensemble of the following five models using WBF[1].

• CenterNet (Backbone:ResNeXt101) + SAHI[2]
• CenterNet (Backbone:ResNet101) + SAHI[2]
• Deformable-DETR (Backbone:ResNet50) + SAHI[2]
• Deformable-DETR (Backbone:ResNet101) + SAHI[2]
• Centripetalnet (Backbone:Hourglass101)

Requirements

Please refer to env.log for the details.

• OS: Ubuntu20.04LTS
• Python: 3.7.16
• CUDA : 11.3
• GPU: Tesla V100-PCIE-32GB x 4
• GCC: gcc (Debian 10.2.1-6) 10.2.1 20210110
• PyTorch: 1.10.1
• TorchVision: 0.11.2
• OpenCV: 4.7.0
• MMCV: 1.6.0
• MMCV Compiler: GCC 9.3
• MMCV CUDA Compiler: 11.3
• MMDetection: 2.24.1+c146527

Installation

1. Create a link file to the data directory

ln -s "path/to/data" data

2. Create a conda environment and activate it.

conda create -n mva python=3.7
conda activate mva

3. Install PyTorch following official instructions

conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge

4. Install MMCV

pip install -U openmim
mim install mmcv-full==1.6.0
pip install -v -e .

Making inference and submission

Inference and submission can be performed by the following command.

bash ./test_submist.sh

In the script, the following steps are performed.

1. Inference on each model with(or without) SAHI
2. Ensemble with WBF
3. Make submission file (./submit/WBF/WBF*.zip)

Training models

Training models can be performed by the following command (but this script is just an example)

bash ./train_test.sh

Training methods differ depending on the model, but the following steps are common.

1. Pre-training with the Drone2021 dataset
2. Finetuning with the mva2023 dataset
3. Extra finetuning with the methods like hard negative training

Please refer to each config file for the details.

References

[1] Weighted Boxes Fusion: Ensembling Boxes for Object Detection Models

[2] SAHI: Self-Adaptive Hierarchical IoU Loss for Small Object Detection