ROADpp_challenge_ICCV2023

ICCV2023 Workshop: The ROAD++ Challenge Track1 Winner (1st)

File Tree

├── demo_pic_and_video
├── ROAD_Waymo_Baseline
├── Road-waymo-dataset
├── runs (Save model's weight)
│   ├── action
│   ├── detect
│   └── location
├── Track1
├── Track2
├── utils

Track1 Full Pipeline

Track2 Full Pipeline

Environment Setup

conda create --name ROADpp python>=3.10
conda activate ROADpp
pip install -r requirement.txt

Pre-train Weight

Download Link (release soon)

├── runs (Save model's weight)
│   ├── action
│   ├── detect
│   └── location

Quick Start

You need to first understand how to configure the YAML file for YOLOv8, as well as the dataset format.

Example for Two branch Track1

cd ROADpp_challenge_ICCV2023
python detect.py --video_path 'xxx' --yolo_path 'xxx' --two_branch True --major_path 'xxx' rare_path 'xxx' --pkl_name 'xxx' --save_res 'xxx'

Config

• mode: Detect mode, only accepts Track1 or Track2.

• video_path: Path to the video.

• yolo_path: Path to the YOLO model.

• two_branch: Indicates whether to use two-branch YOLO.

• major_path: Path to the major YOLO model.

• rare_path: Path to the rare YOLO model.

• devices: GPU number.

• imgsz: YOLO input size.

• video_shape: Original video resolution.

• submit_shape: Final submit shape.

• pkl_name: Submit file name (*.pkl).

• save_res: Save submit file.

• action_detector_path: Path to the action detector model (Track2 only).

• loc_detector_path: Path to the location detector model (Track2 only).

• t2_input_shape: Track 2 input shape.

• windows_size: Sliding windows shape.

ToDo

• Convert Datasets to YOLO format()
• Train YOLOv8 on Track1(train_YOLOv8.py)
• implement Tracklet Function
• Track2 Pipeline
• Two branch Yolo Pipeline
• Implement linear interpolation bbox function
• Complete quick start guide(config using YMAL file)
• Fix T2 interpolation bug
• Two branch T2