bhack

For `nonzero` as `sum` seems already covered in the table at: https://github.com/NVIDIA/cccl/issues/50#issuecomment-1956325564 `cub::DeviceSelect::Flagged` is the only one still needed for large `N`: ```cuda cub::DeviceSelect::Flagged(nullptr, temp_storage_bytes, counting_itr, itr, out_temp.mutable_data_ptr(), (int*)num_nonzeros.get(), N,...

@elstehle We are having another problem related to this with the just release (but popular= model by Meta SAM2: https://github.com/facebookresearch/segment-anything-2/issues/44 Any progress on this? Basically pytorch `nonzero` ops rely on...

Do you have any example of the Online Tapir/BootsTapir inference?

I have also found many of the described problems and false positives matching in these type of sequences. I suppose that for an odometry like camera movement it would also...

> One sanity check is how the model predicts 't 0' points at 256x256 resolution, is it working mostly of the time? I've still not experimented on the same sequence...

> The model is (approximately) estimating whether the result is within an 8 pixel threshold at every resolution; relative to the image size, this threshold gets smaller and smaller later...

Or are you talking about only the "pyramid" effect in iterative approach at section 4 and not about the "feature" backbone pyramid? > 4. Extension to High-Resolution Videos When running...

I also have and extra doubt about section 4. As you are going to work with iterative refinements at 2x iter you are still locked in the "receptive field" of...

@yangyi02 Can this trick used also for training? I want to ask if you also a raw estimate about the memory O of the `build_model_init` feature extractor/backbone vs the tracking...

> Feature backbone is relatively small (similar to ResNet-18), but it's applied across the whole video, so for long videos can be problematic, especially if the video is long. Sorry...