d2.cattle

detectron2 for cattle (boxes, masks, and keypoints)

Installation

conda create -n d2.cattle python=3.8 -y
conda activate d2.cattle 
conda install pytorch==1.10.0 torchvision==0.11.0 cudatoolkit=11.3 -c pytorch

# coco api
pip install pycocotools

# detectron2
python -m pip install detectron2 -f \
  https://dl.fbaipublicfiles.com/detectron2/wheels/cu113/torch1.10/index.html

# cv2
pip install opencv-python

# setuptools
pip install setuptools==59.5.0

Usage

1. Visualize datasets

python visualize_data.py --dataset-name keypoints_train --output-dir data/outtest/viz_back_kp_test/ --source annotation 

2. Training

Example of training keypoints dectection on back cattle dataset, using R50 FPN as backbone.

python train_net.py --config-file configs/CattleKeypoints/keypoints_rcnn_R_50_FPN.yaml

3. Testing

Example of testing keypoints dectection on back cattle dataset, using R50 FPN as backbone.

python train_net.py --num-gpus 1 --config-file configs/cattleKeypoints/keypoints_rcnn_R_50_FPN.yaml --eval-only MODEL.WEIGHTS data/train_outputs/test/model_final.pth OUTPUT_DIR data/train_outputs/test/ 

3.5. Inference

python infer_net.py --num-gpus 1 --config-file configs/CattleKeypoints/keypoints_rcnn_R_50_FPN.yaml  MODEL.WEIGHTS data/train_outputs/test/model_final.pth OUTPUT_DIR data/train_outputs/test/   DATASETS.TEST "('keypoints_test_infer',)"
python infer_net.py --num-gpus 1 --config-file configs/CattleKeypoints/keypoints_rcnn_R_50_FPN.yaml  MODEL.WEIGHTS data/train_outputs/test/model_final.pth OUTPUT_DIR data/train_outputs/train/   DATASETS.TEST "('keypoints_train',)"

For larger batches, paths.json can be loaded with a list of paths to videos desired for predictions. The inference.bash file can then be altered for the number of videos and ran. The predictions for each video will then be stored in "datasets/keypoints/coco_format/annotations/inferences"

4. Demoing

cd demo/
python demo.py --config-file ${config_file_path} \
  --input ${img_path}/*jpg \
  --output ${output_path} \
  --confidence-threshold 0.9 \
  --opts MODEL.WEIGHTS ${model_path}

5. Visualize results

# For set with annotations
python visualize_json_results.py --input /home/ethan/d2.cattle/data/train_outputs/test/inference/coco_instances_results.json --output data/inference/vis --dataset keypoints_test

# For set without annotations
python visualize_json_results_infer.py --input /home/ethan/d2.cattle/data/train_outputs/test/inference/coco_instances_results.json --output data/inference/vis_test --dataset keypoints_test_infer
python visualize_json_results_infer.py --input /home/ptthang/d2.cattle/data/train_outputs/train/inference/coco_instances_results.json --output data/inference/vis_train --dataset keypoints_train

# to merge into videos (with consecutive frames)
python visualize_json_results_infer_merge_video.py --output data/inference/vis_test
python visualize_json_results_infer_merge_video.py --output data/inference/vis_train

New data

• download the json annotation file for each folder
• merge them into one json file, and split into train and test json
• split them into train and test images

The final structure must be

data/
  train_imgs/
    *.jpg
  test_imgs/
    *.jpg
  annotations/
    *.json

Before doing this, you should backup the old data folder. Then run the original script like normal. Or you can create new data, but remember to change all the config and python scripts too.

You can use the notebook make_new_data.ipynb to convert any new data into the required format. Then you should run step 1 to make sure the folder can be visualized correctly. Then you can train it.

visualize

python train_net.py --config-file configs/CattleKeypoints/keypoints_rcnn_R_50_FPN.yaml
python train_net.py --config-file configs/CattleKeypoints/keypoints_rcnn_R_50_FPN_withaligned.yaml


python infer_net.py --num-gpus 1 --config-file configs/CattleKeypoints/keypoints_rcnn_R_50_FPN_withaligned.yaml  MODEL.WEIGHTS data/train_outputs/test/model_final.pth OUTPUT_DIR data/train_outputs/test/   DATASETS.TEST "('keypoints_test_infer',)"
python infer_net.py --num-gpus 1 --config-file configs/CattleKeypoints/keypoints_rcnn_R_50_FPN_withaligned.yaml  MODEL.WEIGHTS data/train_outputs/test/model_final.pth OUTPUT_DIR data/train_outputs/train/   DATASETS.TEST "('keypoints_train',)"

python visualize_json_results_infer.py --input /home/ptthang/d2.cattle/data/train_outputs/test/inference/coco_instances_results.json --output data/inference3/vis_test --dataset keypoints_test_infer
python visualize_json_results_infer.py --input /home/ptthang/d2.cattle/data/train_outputs/train/inference/coco_instances_results.json --output data/inference3/vis_train --dataset keypoints_train


python visualize_json_results_infer_merge_video.py --output data/inference2/vis_test
python visualize_json_results_infer_merge_video.py --output data/inference3/vis_test

Sometimes the model is not confident enough, they will result in empty predictions. In this case, you can try to

• increase the number of training sample and train again
• Make sure the training and testing set are of the same distribution (looks similar)

However, if you still want to force predictions anyway, you can change this field in the configs/CattleKeypoints/keypoints_rcnn_R_50_FPN.yaml

   SCORE_THRESH_TEST: 0.0

By default of detectron2, it is 0.05. By forcing it 0.0, you get low confident prediction too. Maybe we need them, maybe not.

Training data used

Link to folder containing dataset for training with rougly 4600 images, 2300 Thermal, 2300 RGB https://uark-my.sharepoint.com/:f:/g/personal/ejc012_uark_edu/Ev2g78XWrEFPrqXuJahTTScBL7DnSBzUGW0GyDTx4JVFCg?e=fuSZq2

Correction Tool

In the KeyPointCorrector folder is the source code for a windows app used to manually correct data by drag and dropping each point.

Use "browse button" to select the folder you wish to correct. This folder should follow the structure and naming convention

your_folder/ images/ *.jpg metadata_csharp.json

Correct points and then use the "Next" and "Previous" buttons to navigate through each image in the folder. Additionally you can mark an unwanted image for discard by clicking the "Discard" button which will then update the status of the image.

Once you are finished you can click the "Save Button" this will give you the option to select a location for the json file modified from metadata_charp according to your corrections. Additionally, any images marked for discard will be removed from the images folder.

The tool also contains hot-keys to move points quicker. This is done by holing the key asscociated with the desired point and then clicking. This will bring the point to your cursors location. 1-9 correspond to their given numbers while q-r on the keyboard correspond to 10-13.