TileClipper: Lightweight Selection of Regions of Interest from Videos...
summary
TileClipper selects rectangular tiles from H.265-encoded traffic video by using tile bitrate as a noisy proxy for object count, calibrating thresholds by exhaustively searching M/N percentiles over 60 segments to maximize F2. Evaluation on 55 videos (two Indian collections plus benchmarks) compares against Reducto, DDS, static tile removal and CloudSE, reports ≈22% data reduction with 92% detection accuracy (live deployment shows ~87–88% accuracy and >55% bandwidth savings), runs >60 FPS on cheap edge boards like Raspberry Pi 4 and uses YOLOv5 for ground truth.
AI-written summary · about
chapters (10)
- 0:00TileClipper introduction
- 1:06Traffic surveillance pipeline
- 2:51Tile encoding core idea
- 4:38Object detection without DNN
- 7:27Clustering tile selection
- 9:55TileClipper architecture
- 10:39Evaluation setup
- 12:01Performance trade-offs
- 13:04Live deployment results
- 13:46Summary
related talks
24mBeyond the Frame: Using Computer Vision to Unlock the Potential of Video Data
12mCascade models in computer vision to boost accuracy and performance
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
10mLightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...
22mSemantic segmentation of tennis match videos
20mPopup: reconstructing 3D video using particle filtering to aggregate crowd responses
40mSmart Crop and Smart Preview via Video Understanding
15mVideo Analytics with Zero-streaming Cameras