The idea: point an off-the-shelf detector at our sail photos with the plain-English labels "boom" and "batten pocket" — no training, no fine-tuning — and see if it can box those parts on its own. If it can, that's a free signal to help SAM3 draw sail outlines or catch a stripe line that clearly isn't on the sail. This was meant to be a quick 10-photo eyeball check, not a real build.
A previous attempt at this exact test caused the detector's Python process to balloon to 6.5GB of memory and lock up the machine. This session checked before touching anything: none of the required software (mmdet, mmcv, mmengine) is installed, no model weights are cached anywhere on disk, and the piece that caused the crash last time — the model itself running in memory — is a known 6.5GB-plus load, not a small download. Rather than risk repeating that crash, this session stopped and confirmed the safe path is a cloud GPU pod, the same kind already used for other model spikes in this project.
| Check | Result |
|---|---|
| Model weights cached locally | No — nothing found anywhere on disk |
| Required software installed | No — mmdet/mmcv/mmengine/transformers all missing from the inference environment |
| Forward passes run | 0 |
| Photos screened | 0 of the planned ~10 |
| Smallest checkpoint found (HF transformers variant, "tiny") | ~657 MB on disk |
| Checkpoint for the actual MM-Grounding-DINO-Tiny model this test is named for | ~916 MB on disk (needs mmdet/mmcv, not installed) |
| Known runtime memory need (from the prior crash on this same test) | 6.5 GB — this is the real blocker, not the download size |
| Machine put at risk this session | No — only checked file sizes over the network, no model was ever loaded |
Any cloud GPU box (nothing exotic — the same class used for prior SAM3 spikes here), the detector loaded with the "boom" / "batten pocket" text prompts, and about 10–20 photos from the frozen holdout set streamed one at a time. Under an hour of GPU time. Someone would then eyeball whether the boxes land on the real boom and battens and report a hit rate — that's the actual test, still not run.
Source: model-mmgdino.json.