Closed lvoursl closed 2 months ago
In this PR we're adding updated version of CoTracker2 (named CoTracker2.1) which follows the same architecture but was trained with different window size (16 instead 8) that leads to better performance on TAP-Vid benchmarks.
The inference code via torch.hub is updated and pushed, new metrics are also release in the query first mode.
@lvoursl Hi, would version 2.1 conduct a training cost and inference speed comparison with version 2.0?
Hi @chenshihfang, we haven't done in-depth benchmarking but our empirical observation that bigger window size leads to faster training and inference.
In this PR we're adding updated version of CoTracker2 (named CoTracker2.1) which follows the same architecture but was trained with different window size (16 instead 8) that leads to better performance on TAP-Vid benchmarks.
The inference code via torch.hub is updated and pushed, new metrics are also release in the query first mode.