Dynamic Curriculum Learning for Great Ape Detection in the Wild

Xinyu Yang, Tilo Burghardt, Majid Mirmehdi

This repository is the official implementation of paper Dynamic Curriculum Learning for Great Ape Detection in the Wild.

Installation

Requirements

• Linux, CUDA>=9.2, GCC>=5.4

• Python>=3.7

Install Dependencies

1. create a conda environment:

conda create -n DCLNet python=3.7 pip

2. activate the environment:

conda activate DCLNet

3. Installation pytorch follow the official doc install at here

conda install pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=10.2 -c pytorch

4. Installation other packages

pip install -r requirements.txt

5. Install CUDA extension

cd ./models/ops
sh ./make.sh

6. test your CUDA extension

python test.py

Test Environments

CPU: AMD EPYC 7543 (128) @ 2.794GHz
GPU: NVIDIA GRID A100X
OS: CentOS Linux release 7.9.2009 (Core) x86_64
nvcc version: 11.2
GCC version: 7.5.0
python version: 3.7.11

Usage

Dataset preparation

MSCoco

Please download COCO 2017 dataset and organize it in the following structure:

data/
└── coco
    ├── test2017
    ├── train2017
    ├── val2017
    └── annotations
        ├── instances_train2017.json
        ├── instances_val2017.json
        └── semi_supervised
            ├── instances_train2017.1@10.json
            ├── instances_train2017.1@10-unlabeled.json
            ├── instances_train2017.1@1.json
            ├── instances_train2017.1@1-unlabeled.json
            ├── instances_train2017.1@5.json
            ├── instances_train2017.1@5-unlabeled.json
            ├── instances_train2017.2@10.json
            ├── instances_train2017.2@10-unlabeled.json
            ├── instances_train2017.2@1.json
            ├── instances_train2017.2@1-unlabeled.json
            ├── instances_train2017.2@5.json
            ├── instances_train2017.2@5-unlabeled.json
            ├── instances_train2017.3@10.json
            ├── instances_train2017.3@10-unlabeled.json
            ├── instances_train2017.3@1.json
            ├── instances_train2017.3@1-unlabeled.json
            ├── instances_train2017.3@5.json
            ├── instances_train2017.3@5-unlabeled.json
            ├── instances_train2017.4@10.json
            ├── instances_train2017.4@10-unlabeled.json
            ├── instances_train2017.4@1.json
            ├── instances_train2017.4@1-unlabeled.json
            ├── instances_train2017.4@5.json
            ├── instances_train2017.4@5-unlabeled.json
            ├── instances_train2017.5@10.json
            ├── instances_train2017.5@10-unlabeled.json
            ├── instances_train2017.5@1.json
            ├── instances_train2017.5@1-unlabeled.json
            ├── instances_train2017.5@5.json
            └── instances_train2017.5@5-unlabeled.json

The partial labelled data (PLD) are split by:

cd ./datasets/prepare_coco
sh ./run.sh

PanAfrican

Coming soon! Stay tuned !

Semi-supervised Mix-training

COCO training

1. running with 1% PLD at fold1 (4 high-end GPUs required):

sh ./train_coco_1percent_1fold.sh

2. running with 5% PLD at fold1 (4 high-end GPUs required):

sh ./train_coco_5percent_1fold.sh

2. running with 10% PLD at fold1 (4 high-end GPUs required):

sh ./train_coco_10percent_1fold.sh

PanAfrican training

Coming soon! Stay tuned

Models checkpoints

Setting	Dataset	mAP	mAP50	map75	checkpoint (Student and Teacher)
1% PLD	COCO	17.34	31.00	17.35	ckpt
5% PLD	COCO	29.75	46.68	31.77	ckpt
10% PLD	COCO	34.45	51.93	37.14	ckpt
10% PLD	PanAfrican	45.96	78.10	47.67	coming soon
20% PLD	PanAfrican	59.01	89.23	66.95	coming soon
50% PLD	PanAfrican	63.39	92.96	70.00	coming soon

Citation

If you are considering using this codebase, please cite our work:

@misc{yang2021dcldet,
      title={Dynamic Curriculum Learning for Great Ape Detection in the Wild},
      author={Xinyu Yang and Tilo Burghardt and Majid Mirmehdi},
      year={2022},
      eprint={2205.00275},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Credits

This repository builds on previous works codebase 1. Deformable DETR 2. DETReg. Please consider citing these works as well.