Skip to content

README.md

Latest commit

 

History

History
25 lines (20 loc) · 1.88 KB

README.md

File metadata and controls

25 lines (20 loc) · 1.88 KB

StrongSORT Usage

Installation

This part is based on the StrongSORT project. Please first check the installation steps on the StrongSORT project to build the conda environment. We only use the DeepSORT method of the project.

Overview

After we get the inference result (instance segmentaion result) of the main model, we need to generate the object trajectories. It includes following steps.

  1. Convert coco format instance segmentation results to txt file.
  2. Convert ReID results to box and features needed by StrongSORT.
  3. Use the DeepSORT method implemented in the StrongSORT project to link the same objects across frames. This step is termed Data Association.
  4. Match masks of the instance segmentation result to boxes in the tracks of the data association result.
  5. Combine results of different classes together, and remove overlapping pixels. Overlapping pixels are not allowed in the MOTS task. This step generates the final MOTS results, which are in fact mask-based trajectories.

Then you can do the evaluation with the MOTS results.

Usage

We provide two different ways to use this code, described below.

One program per function

Programs for different functions are in the my_code folder for KITTI. Check this description to understand the function of each python program.

For BDD dataset, check the my_code_bdd folder, and its description.

One script for all functions

Use the script to finish all 5 steps with one operation for KITTI. It call python programs in the my_code_pipeline folder. Ensure these paths in ths script are correct on your local device.

For BDD dataset, check this script and the my_code_pipeline_bdd folder.