Trinetra: O(N) Video Ingest, O(1) Query — A Proactive Temporal Event Graph for Long-Form Video Understanding

Abstract - Current video understanding systems are fundamentally reactive: they re-process every frame at query time, incurring quadratic O (T) attention complexity that caps practical span at 30–60 minutes and costs 15–24 per query on hour-long videos. We argue this is an architectural choice, not a physical law. We introduce Trinetra (ThiriNay-thra, Tamil for “three-eyed, ” evoking simultaneous perception across temporal scales) with a provably better total-cost model: O (T) + O (Q) versus reactive systems’ O (Q · T) for Q queries over a video of length T. The core of Trinetra is a Temporal Event Graph (TEG): a hierarchical representation built once during ingest and queried via dual FAISS indices without ever re-processing the video. Our Multi-Scale Temporal Attention Shift (TAS) performs a wavelet-like multiresolution decomposition of the video embedding stream — simultaneously capturing gestures (highfrequency), actions (mid-frequency), and scenes (low-frequency) — combined with a discrete firstorder temporal derivative ∆t = et − et−1 that detects causal transitions in O (1) per frame. The TEG hierarchy (Frame → Event → Chapter → Video) constitutes a semantic compression ladder: information-theoretic entropy decreases monotonically up the hierarchy, retaining only causally relevant structure. Eight temporal memory tokens provide O (1) -space low-rank compression of unbounded video history. Evaluated on a Raspberry Pi 5 assembly tutorial (1, 236 s) on a consumer RTX 3050 (4 GB VRAM): 60% category success rate without audio (estimated 80%+ with audio enabled), 25. 8 s ingest, 15. 2 s per query, 317, 500× fewer query-time FLOPs than commercial VLMs, at under 0. 01 per 100 queries versus 1, 500–2, 430 for GPT-4o / Gemini / ClaudeCopyright - @S KhavinIs supplement to - https: //doi. org/10. 5281/zenodo. 18850768 This Novel work presents a self-driven research effort aimed at advancing problem of long video understanding in computer vision, with the goal of submission to CVPR. Developed independently by a 20y old researcher and without formal institutional backing. The current version may have minor formatting limitations, but the underlying ideas, methodology, and experimental direction are actively being refined and extended. The focus is on building non-trivial, high-impact approaches rather than incremental improvements. I am actively seeking research internships and collaborations where I can contribute meaningfully while working alongside experienced researchers. I bring a high level of initiative, the ability to move quickly from idea to implementation, and a strong commitment to rigorous problem-solving. If you are interested in the work, open to collaboration, or willing to provide critical feedback, feel free to reach out: Contact: skhavin. res@gmail. com