Big day.

The research paper I've been building towards for the past year just dropped as a preprint with a DOI. Submitted to a Springer journal, currently in peer review.

https://dx.doi.org/10.21203/rs.3.rs-9552341/v1

The project is GlyphMotion — a real-time 4K multi-object tracking pipeline. You feed it video, it tracks every object across every frame, and outputs annotated 4K with audio intact. The interesting engineering is in how it gets there.

Synchronous pipelines can't handle this workload. We built a three-thread async architecture — reader, inference, writer — with bounded queues between them. Jitter went from 337ms down to 27ms on modern hardware, and from 21 seconds down to 46ms on older hardware.

The finding that caught us off guard: raw 4K footage is actually bad for tracking. The sensor noise in uncompressed footage constantly throws off the tracker's identity associations. Compressing at CRF 24 before running inference stabilised it — MOTA jumped from 55 to 85. Compression improving accuracy is not intuitive, but the data is consistent across 159 videos.

The quality loss from compression is recovered through a layer we call HFDR — High-Frequency Detail Reinjection — which adds back the spatial detail that compression strips. Final average VMAF across 159 videos: 96.67.

Built with Sayan Sarkar @sayann70. Supervised by Dr. Kretika Goel, DIT University. We started this right after class 12 boards, before college.

Read the preprint if you're curious. Happy to answer questions.