Updating a dataset of labelled objects on raw video sequences with unique object IDs

Area of Science:

• Computer Vision
• Video Processing
• Machine Learning

Background:

• Existing datasets like SFU-HW-Objects-v1 lack object tracking capabilities.
• Evaluating object tracking performance under video compression requires specialized datasets.
• High Efficiency Video Coding (HEVC) Common Test Conditions (CTC) are standard benchmarks.

Purpose of the Study:

• Introduce SFU-HW-Tracks-v1, an enhanced dataset for object tracking.
• Provide unique object identities (IDs) for tracking analysis.
• Facilitate research on the interplay between video compression and object tracking.

Main Methods:

• Annotated uncompressed video sequences from HEVC CTC.
• Included ground truth for object class ID, object ID, and bounding boxes per frame.
• Ensured unique object IDs for consistent tracking across frames.

Main Results:

• SFU-HW-Tracks-v1 dataset is now available with detailed object tracking annotations.
• The dataset supports evaluation of object tracking algorithms on HEVC sequences.
• Enables quantitative study of compression's impact on tracking accuracy.

Conclusions:

• SFU-HW-Tracks-v1 addresses the need for object tracking datasets in video compression research.
• The dataset will advance the understanding of tracking performance degradation due to compression.
• Facilitates development of more robust object tracking methods for compressed video.