A Controlled Dataset to Evaluate Temporal Sensitivity in BDQ Encoding

Project Overview

This repository provides the code and data pipeline for constructing a control dataset designed to evaluate the temporal sensitivity of the BDQ encoder introduced in the paper Privacy-Preserving Action Recognition via Motion Difference Quantization [1].

The goal is to test whether BDQ's performance in action recognition and identity suppression truly relies on inter-frame motion cues. Based on a subset of the IXMAS dataset [2], we construct two matched conditions for each video clip:

• Original sequences: 32 consecutive frames with intact temporal order
• Shuffled sequences: the same frames divided into 4 chunks of 8 and randomly reordered to disrupt global temporal flow

Directory Structure

Datasets & Metadata

• Raw datasets available from IXMAS Actions – New Views and Occlusions - CVLAB - EPFL
• datasets/ixmas_clips_6.json: structured metadata such as action and subject ID for selected clips

Preprocessing Scripts

• preprocess/IXMAS_720: 720 manually selected representative frames (10 subject $\times$ 12 classes $\times$ 2 viewpoints $\times$ 3 takes)
• preprocess/ixmas_extract_frame.py: extracts and saves a representative frame from each IXMAS video
• preprocess/ixmas_extract_vid.py: copies videos matching selected frame names (e.g., from IXMAS_720/)
• preprocess/ixmas_extract_vid_class.py: filters and saves videos of the six selected action classes
• preprocess/ixmas_parser.py: parses video filenames into structured metadata

Control Dataset Construction

1. Download and organize raw data Download the original IXMAS ROIs and place it under the preprocess/IXMAS_raw/.
2. Install dependencies:

pip install -r requirements.txt

1. Preprocess original dataset:

# Optional: extract one representative frame from each video
# python preprocess/ixmas_extract_frame.py

python preprocess/ixmas_extract_vid.py # Extract all videos from chosen frames 
python preprocess/ixmas_extract_vid_6.py # Keep only the 6 selected actions
python preprocess/ixmas_parser.py # Generate structured metadata 

1. Generate the control dataset:

python trim_vid.py # Cut videos to exactly 32 consecutive frames 
python shuffle_chunks.py # Shuffle each video 

References

[1] S. Kumawat and H. Nagahara, “Privacy-Preserving Action Recognition via Motion Difference Quantization,” Aug. 2022.

[2] D. Weinland, R. Ronfard, and E. Boyer, “Free viewpoint action recognition using motion history volumes,” Computer Vision and Image Understanding, vol. 104, pp. 249–257, Nov. 2006.