CVWC2019-Amur-Tiger-Detection

Example output of Faster_Rcnn-R-50-FPN_1x .

⭐This code is for the Tiger Re-ID in the Wild track (detection part) CVWC2019 @ICCV19 Workshop:

To generate detected tiger pictures for Tiger Re-ID in the Wild.

---

🏃Getting Start

1️⃣ Clone the repo:

git clone https://github.com/LcenArthas/CVWC2019-Amur-Tiger-Detection.git

2️⃣ Dependencies

Tested under python3. Ubantu16.04

• python packages
  • pytorch==0.4.1(Note: V1.0.1 may result in an error)
  • torchvision>=0.2.0
  • cython==0.29.3
  • cffi==1.12.3
  • matplotlib==3.1.1
  • tqdm==4.32.2
  • numpy==1.16.4
  • scipy==1.2.1
  • opencv==4.1.0.25
  • pyyaml==5.1.1
  • packaging==19.0
  • pycocotools — for COCO dataset, also available from pip.
  • tensorboardX — for logging the losses in Tensorboard
• An NVIDAI GPU and CUDA 8.0 or higher. Some operations only have gpu implementation.

3️⃣ Compilation

Compile the CUDA code:

cd lib 
sh make.sh

It will compile all the modules you need, including NMS, ROI_Pooing, ROI_Crop and ROI_Align. (Actually gpu nms is never used ...)

---

🏃 Train

1️⃣ Data Prearation

• Train Dataset

🔸 Creat a new folder named /Original_train/ under the {repo_root}/data/:

cd data
mkdir Original_train

🔸 Put the train dataset(two folder: atrw_anno_detection_train & atrw_detection_train) in the {repo_root}/data/Original_train/ folder under the repo.

🔸 Transform the data style for the model

python make_coco_data.py

2️⃣ Pre-trained weight

🔸 Creat a new folder named /pretrained_model/ under the {repo_root}/data/:

cd data
mkdir pretrained_model

🔸 Download the pre-trained weighte and put it(resent50_caffe.pth) into the {repo_root}/data/pretrained_model/.

• Pre-trained weight

And make sure the repo files as the following structure:

{repo_root}
├── configs
├── demo
├── lib
├── data
|   ├── coco
│   │    ├── anntations 
│   │    |    └── instances_train2017.json
│   │    └── images
│   │         └── train2017
│   │               ├── 0000.jpg
│   │               ├── 0002.jpg
│   │               ├── 0003.jpg
│   │               ├── 0004.jpg
│   │               ├── 0005.jpg
│   │               └── ...
|   ├── Original_train
|   |    ├── atrw_ann_detection_train
|   |    ├── atrw_detection_train
|   |    └── data
|   └── pretrained_model
|        ├── resnet50_caffe.pth
|        └── resnet101_caffe.pth  
├── tools
└── make_coco_data.py
    

👏 Train Now!

cd tools
python train_net_step.py

Eventually the trained model will be saved in {repo_root}/tools/Outputs/

---

🏃 Inference

1️⃣ Data Preparation

🔸 Creat a new folder named /test/ under the {repo_root}:

mkdir test

🔸 Put the test images in the {repo_root}/test/ folder under the repo.

• Test Dataset

2️⃣ Download Pretrained Model

I use Faster-rcnn-Resnet50-FPN to train my model.

• Trained weight

🔸 Download it and create a new folder under the {repo_root} named /trained_weight/

mkdir trained_weight

🔸 Put the trained weight(best_model.pth) into the {repo_root}/trained_weight/.

And make sure the repo files as the following structure:

{repo_root}
├── configs
├── demo
├── lib
├── test
|   ├── 0001.jpg
│   ├── 0002.jpg
│   ├── 0003.jpg
│   ├── 0004.jpg
│   ├── 0005.jpg
│   └── ...
├── tools
├── trained_weight
│   ├── best_model.pth
└── make_coco_data.py
    

👏 Inference Now!

cd tools
python infer_simple.py

This process will take about 15 minutes, just a moment, please.

---

Run this scrip will generate 3 files in the {repo_root/}:

• det_submission.json — for the Tiger Detection track, you can submit in the Tiger Detection track (0.45988 mAP in the Public Leaderboard).

• wide_box.json — for the Tiger Re-ID in the Wild track.

• reid_test(a folder) — for the Tiger Re-ID in the Wild track, it contains images that have been detected and croped.