6 posts
A uniform "every Nth frame" sampler wastes labeler time on near-identical frames the model already nailed. Four signals do better — segment boundaries, per-segment uniform, low confidence, bbox jumps.
`mediapipe`, `ultralytics`, `cv2` are slow to import and need model weights at runtime. The trick that keeps the test suite small, fast, and weights-free is putting those imports inside function bodies, not at module top level.
When you harvest Label Studio exports to fine-tune the next model, treating *everything* in the export as ground truth is how you train the model on its own predictions. The filters worth applying before a single byte goes into a training set.
A deep dive on the multi-project pattern for video pre-annotation — what forces the split, how one episode fans out, and when not to fight Label Studio's data model.
Inference frame rate and review-frame sampling look like one thing and aren't. What each knob actually buys, and what breaks if you treat them as the same.
A Phase-1 pipeline for embodied-robot video data — MediaPipe + YOLO inference, action segmentation, Label Studio import — plus the boring path-abstraction decision that kept it from collapsing.