Right now, with our 2026 model, we use non-overlapping segments during inference. benchmark running with a small overlap, then aggregating the prediction - so like, 50 frames overlap, or maybe 100, or 200 (worth trying)
This is important, since it gives the model more context about signs in the beginning/end.
Right now, with our 2026 model, we use non-overlapping segments during inference. benchmark running with a small overlap, then aggregating the prediction - so like, 50 frames overlap, or maybe 100, or 200 (worth trying)
This is important, since it gives the model more context about signs in the beginning/end.