Post
128
So I built a multimodal video annotation pipeline in my spare time, as you do.
corpus-mill turns any long-form video with people on camera into a time-aligned event corpus across audio, vision, OCR, faces, brand observations, music, and clip-worthy moments. Runs entirely on local GPU because — and I cannot stress this enough — your footage has no business being on someone else's servers.
The honest origin: I needed real multimodal supervision data, the public corpora are weirdly thin once you need per-frame / per-speaker / per-second labels with provenance, so I built one. Then it grew. Then I looked up and it was 30K LOC and ~30 stages and I thought, ok, maybe other people would want this.
Stack is the usual suspects: Whisper-large-v3 (faster-whisper), pyannote-3.1 (which secretly drags in 433 NeMo modules — surprise!), Qwen2.5-VL-7B for vision/OCR/shoppable detection, dlib + YuNet for faces, qwen2.5:7b / qwen3:14b via local Ollama for the LLM passes, chromaprint + PDQ for fingerprinting. Outputs as Parquet + SQLite. Apache 2.0.
There's a Docker compose that works, after I spent a day discovering that faster-whisper wants CUDA 12 cuBLAS while pyannote 4 wants CUDA 13, and the answer is "install both, point LD_LIBRARY_PATH at the cu12 wheels, ship it." That's now baked in. You're welcome.
Spare-time project, bugs are real, fixing them for your specific footage is on you. If you're training multimodal models and want a corpus pipeline you fully control on-prem, this might save you months. If not, the README is at least mildly entertaining.
https://github.com/cahlen/corpus-mill
corpus-mill turns any long-form video with people on camera into a time-aligned event corpus across audio, vision, OCR, faces, brand observations, music, and clip-worthy moments. Runs entirely on local GPU because — and I cannot stress this enough — your footage has no business being on someone else's servers.
The honest origin: I needed real multimodal supervision data, the public corpora are weirdly thin once you need per-frame / per-speaker / per-second labels with provenance, so I built one. Then it grew. Then I looked up and it was 30K LOC and ~30 stages and I thought, ok, maybe other people would want this.
Stack is the usual suspects: Whisper-large-v3 (faster-whisper), pyannote-3.1 (which secretly drags in 433 NeMo modules — surprise!), Qwen2.5-VL-7B for vision/OCR/shoppable detection, dlib + YuNet for faces, qwen2.5:7b / qwen3:14b via local Ollama for the LLM passes, chromaprint + PDQ for fingerprinting. Outputs as Parquet + SQLite. Apache 2.0.
There's a Docker compose that works, after I spent a day discovering that faster-whisper wants CUDA 12 cuBLAS while pyannote 4 wants CUDA 13, and the answer is "install both, point LD_LIBRARY_PATH at the cu12 wheels, ship it." That's now baked in. You're welcome.
Spare-time project, bugs are real, fixing them for your specific footage is on you. If you're training multimodal models and want a corpus pipeline you fully control on-prem, this might save you months. If not, the README is at least mildly entertaining.
https://github.com/cahlen/corpus-mill
