Image Quality Analysis

Evaluating the impact of compression, noise, and artifacts on Deep Learning models.

Most Computer Vision models are trained on “pristine” datasets (like COCO or ImageNet), but real-world surveillance involves transmission lag, heavy H.264/H.265 compression, and sensor noise. This research pillar focuses on quantifying exactly how much “intelligence” we lose when video quality drops.

Core Research Themes

1. Object Detection under Compression

We evaluated state-of-the-art detectors (YOLOv3, Faster R-CNN, SSD) against varying levels of bit-rate compression. Our findings highlight a critical “cliff” where mAP (mean Average Precision) drops significantly even before the artifacts are highly visible to the human eye.

2. Background Subtraction & Temporal Artifacts

Surveillance often relies on Background Subtraction (BGS) to trigger alerts. We analyze how GOP (Group of Pictures) structures and motion estimation artifacts in compressed video create “ghosting” effects, leading to false positives in abnormality detection pipelines.

Key Publications

This project encompasses our systematic evaluation of 4 SOTA deep neural network models, showing that existing detectors are highly susceptible to quality distortions stemming from video acquisition.

Visualizing how compression artifacts lead to missed detections and bounding box jitter in high-density urban scenes.
Read Paper (VISAPP)

References

2023

  1. Image Quality Assessment for Object Detection Performance in Surveillance Videos.
    Poonam Beniwal, Pranav Mantini, and Shishir Shah
    In VISIGRAPP (4: VISAPP), 2023

2022

  1. Image Quality Assessment using Deep Features for Object Detection.
    Poonam Beniwal, Pranav Mantini, and Shishir K Shah
    In VISIGRAPP (4: VISAPP), 2022

2019

  1. Understanding How Video Quality Affects Object Detection Algorithms.
    Miloud Aqqa, Pranav Mantini, and Shishir K Shah
    In VISIGRAPP (5: VISAPP), 2019