PMMA dataset processing and training document

Data Link:

PMMA Dataset (via DOI): https://doi.org/10.5683/SP3/XJPQUG

Data preparation

Download original videos from link and put it in 'data_processing/raw_data/original_videos'

1. Go to the data path

   cd data_processing
   

2. Convert videos to images

   python step_1_video_2_images.py
   

3. Merge images for building training/validation/test sets

   python step_2_merge_images.py
   

4. Convert CVAT annotations to COCO format

   python step_3_convert_cvat_to_coco_full.py
   

5. Split the images into training/validation/test sets

   python step_4_splitting.py
   

Training

MMDetection is used to train Faster R-CNN, DINO, CenterNet, YOLOX, DETR, and Deformable DETR. Please refer to the MMdetection documentation for environment setup and installation instructions.

RT-DETR is used to train RT-DETR. Please refer to the RT-DETR documentation for environment setup and installation instructions.

cd training

MMdetection:

Exmaple for training:

1. Go to the MMdetection path

   cd /path/to/MMdetection
   

2. Use Faster R-CNN as an example, train the model:

   CUDA_VISIBLE_DEVICES=0,1,2,3 ./mmdetection/tools/dist_train.sh configs_training/configs_faster_rcnn.py 4
   

   where 0,1,2,3 are available GPU names, 4 is the number of GPUs

RT-DETR:

1. Copy the data yaml and model info to RT-DETR

   cp configs_training/coco_detection_PMMA.yml /path/to/RT-DETR/redetrv2-pytorch/configs/dataset
   
   

   cp configs_training/rtdetrv2_r50vd_6x_coco_run.yml /path/to/RT-DETR/redetrv2-pytorch/configs/redetrv2
   

2. Go to RT-DETR path:

   cd /path/to/RT-DETR/redetrv2-pytorch/
   

3. Train the model:

   python tools/train.py -c configs/rtdetrv2/rtdetrv2_r50vd_6x_coco_run.yml
   

Dataset Layout

PMMA/
├── raw_annotations/
│   ├── video_1_annotations.xml
│   └── video_2_annotations.xml
│   └── video_3_annotations.xml
│
├── annotations/
│   ├── train.json
│   └── val.json
│   └── val.json
│
├── images/
│   └── video_2/
│         ├── video_2_1.zip
│         ├── video_2_2.zip
│         ├── video_2_3.zip
│         ├── video_2_4.zip
│         ├── video_2_5.zip
│         ├── video_2_6.zip
│
├── videos/
│   ├── video_1.mp4
│   ├── video_3_1.mp4
│   ├── video_3_1.mp4
└── README.md

Notes: The original video_2 is not provided as video_1 and video_3, as the video was involved with people who are volunteers. Instead, images are provided in "images/video_2".