v0.4.0: video understanding

Understand and search video: shots and keyframes, topics per shot and per video, faces grouped into people with time intervals, animals and food on a timeline, speech to text, a summary in English and Persian, a self-contained HTML report, and search across videos by text, speech or face.

Tested on real footage with the real models. What was and was not measured.

vie video analyze clip.mp4 --html report.html
vie video search --query "a goal being scored"

v0.3.0: everything connected and measured

A full audit found 25 defects behind 198 passing tests, and all were fixed. Every upgrade now reaches the running system, and thresholds are measured: on 72 public images with the default model, species precision went from 57% to 94% and food from 46% to 92%, at 100% recall.

Also adds batched indexing, optional FAISS, schema migrations, lazy model loading, vie doctor, HTML search reports, and a CI job that runs every real model.

Existing indexes must be rebuilt to benefit: vie index --rebuild. See the audit and the changelog.

v0.2.0: SigLIP

Moved to SigLIP so400m with a hybrid species path. Superseded: the audit in v0.3.0 found that species search crashes in this version and that several upgrades were not connected. Use v0.3.0 or later.

v0.1.0: first complete package

Notebooks extracted into an installable package with a CLI, config, tests and CI, on CLIP ViT-L/14.

v0.0.1: original artefacts

The earliest state in git: the roadmap, the original demo screenshots and the original architecture diagram. The original Colab notebooks' code was never committed, so this is a record of the project's starting point and does not run.