Research map / Detection and segmentation / Tracking
Descriptor-In-Pixel : Point-Feature Tracking For Pixel Processor Arrays
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Descriptor-In-Pixel, an in-sensor point-feature detection and tracking method on the SCAMP-7 pixel processor array that runs at over 3000 FPS.
Problem
Point-feature detection and tracking on conventional cameras requires transferring whole images to external processors each frame, creating a latency and bandwidth bottleneck for high-speed, low-power applications.
Approach
Store an 8-bit binary BRIEF-like descriptor in each pixel-processor and compute descriptor responses for all PEs in parallel; a patchwork response map (tracked-feature patches plus a randomized search descriptor elsewhere) is used for both detection via blob detection and tracking via local maxima/NMS, so only sparse feature locations and descriptors leave the sensor.
Key results
- Implementation on SCAMP-7 PPA runs at over 3000 FPS and tracks point-features reliably under violent motion.
- Sensor output reduced by greater than 1000x versus raw image output, averaging about 2 bytes per feature per frame.
- Under the Translate motion, about 20% of features tracked by the proposed method survive over 15 seconds versus 0% for FAST-based tracking on SCAMP-7 at 1000 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Descriptor-In-Pixel tracking instead of matching FAST corner points by proximity | SCAMP-7 feature tracking · Translate motion | feature lifetime: 90% ... lost within 5 seconds; 0% ... tracked for over 15 seconds; around 20% ... tracked for over 15 seconds | Helped |
Surprising findings
- Hamming-distance-based descriptor matching can produce unreliable response maps for tracking, motivating the analogue pixel-pair weighted response instead.
Limitations
- Tracking is not invariant to rotation or scale because only location and descriptor are stored in-pixel.
- Only 8-bit descriptors fit in the limited per-PE memory of SCAMP-7.
- Evaluation is limited to a comparison with a FAST-based baseline due to lack of high-frame-rate datasets and difficulty uploading frames to the PPA.
Experimental setup
Benchmarks: SCAMP-7 feature lifetime comparison vs FAST tracking under Translate and other sensor motions
Metrics: feature lifetime distribution processing speed (FPS) compute time per frame (µs) data reduction ratio
Baselines: FAST-based tracking on SCAMP-7 BITVO
Builds on: BRIEF BRISK FREAK FAST ORB
Closest papers by meaning
- PointPillars: Fast Encoders for Object Detection From Point Clouds (CVPR 2019 · similarity 0.84)
- TTAPFormer: Robust Arbitrary Point Tracking via Transient Asynchronous Fusion of Frames and Events (CVPR 2026 · similarity 0.83)
- Dense Optical Tracking: Connecting the Dots (CVPR 2024 · similarity 0.82)
- Tracking Everything Everywhere across Multiple Cameras (AAAI 2025 · similarity 0.82)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · similarity 0.81)
- PTT: Point-Trajectory Transformer for Efficient Temporal 3D Object Detection (CVPR 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.