Skip to content

Latest commit

 

History

History
100 lines (79 loc) · 5.08 KB

README.md

File metadata and controls

100 lines (79 loc) · 5.08 KB

-----------------------------------------------------

➤ Week 3

During the third week of our project, we focused on estimating the optical flow of a video sequence, estimating the optical flow and trying to improve an object tracking algorithm and finally we provided results on data from the CVPR AI City Challenge.

-----------------------------------------------------

➤ Available tasks

  • Task 1: Optical Flow
    • Task 1.1: Optical Flow with Block Matching
    • Task 1.2: Off-the-shelf Optical Flow
    • Task 1.3: Object Tracking with Optical Flow
  • Task 2: Multi-target single-camera (MTSC) tracking
    • Task 2.1: SEQ 01
    • Task 2.2: SEQ 03
    • Task 2.3: SEQ 04

-----------------------------------------------------

➤ Usage

Task 1.1

In this task we compute the optical flow using the block matching method. The main objective of the task can be performed using script task1_1.py, which can be run the following way:

  python task1_1.py [--sequence SEQUENCE] [--block-size BLOCK_SIZE] [--search-area SEARCH_AREA] [--metric METRIC] [--direction DIRECTION] [--max-level MAX_LEVEL]

Where --sequence is the sequence that we want to compute the optical flow from. By default, the sequence is 45. The rest of the input parameters are the following:

  • --block-size is the size of the block used for the block matching method. By default, the block size is 16.
  • --search-area is the size of the search area used for the block matching method. By default, the search area is 20.
  • --metric is the metric used for the block matching method. By default, the metric is sad (Sum of Absolute Differences). Can also be ssd(Sum of Squared Differences) and ncc(Normalized Cross Correlation Coefficient).
  • --direction is the direction of the optical flow. By default, the direction is forward.
  • --max-level is the maximum level of the pyramid used for the block matching method. By default, the maximum level is 3.

Task 1.2

The main objective of the task can be performed using script task1_2.py, which can be run the following way:

  python task1_2 [--sequence SEQUENCE]

Where sequence is the sequence that we want to compute the optical flow from. By default, the sequence is 000045.

Additionally, there is another script, that generates a video of the optical flow generated from a given input video. It can be run the following way:

python optical_flow_video.py [--flow-method FLOW_METHOD] [--input-path INPUT_PATH] [--output-path OUTPUT_PATH]

Where:

  • --flow-method is the method used (either pyflow or farneback)
  • --input-path is the input path for the given video
  • --output-path is the output path for the generated video

Task 1.3

This task can be executed with the following script:

python task1_3.py [--predictions-file PREDICTIONS_FILE] [--tracking-file TRACKING_FILE] [--video-path VIDEO_PATH]
                  [--visualization-path VISUALIZATION_PATH] [--visualize-video VISUALIZE_VIDEO] [--flow-method FLOW_METHOD]
                  [--bbx-flow-method BBX_FLOW_METHOD]

Where:

  • --predictions-file is the name of prediction json file (YOLO style)
  • --tracking-file is the name of output file
  • --video-path is the path to video for visualization
  • --visualization-path is the path to save visualization video
  • --visualize-video is a bool to visualize video on execution
  • --flow-method is the optical flow method
  • --bbx-flow-method is the method to shift the bounding boxes from the optical flow

Disclaimer: this scripts runs tracking given a json object detection prediction file with YOLO style.

Task 2

The set-up for the TrackEval benchmark can be sped up through the use of the following scripts.

First run

python createtrack1folders.py

to create the gt aicity19 benchmarks with the appropiate ground truth folders for each sequence.

Then create the challenge folder aicity19-train. The following script creates the tracker folder and populates it with the prediction files outputted by the task 1.3 script

python convert_to_trackeval.py

Warning: json files are expected at the directory Week3/trackings

Finally run the evaluation

python task2.py --GT_FOLDER .\TrackEval\data\gt\mot_challenge\ --TRACKERS_FOLDER .\TrackEval\data\trackers\mot_challenge\ --BENCHMARK aicity19 --METRICS HOTA Identity --DO_PREPROC False --TRACKERS_TO_EVAL farnebackmedian

-----------------------------------------------------

➤ Requirements

Install the requirements with the following command:

pip install -r Week3/requirements.txt