tomztyang

That inference time is tested on a TITAN V GPU card with some optimization like merging BN, parallel computation in the first SA layers. These optimizations are mentioned in our...

@tianweiy , On the NuScenes dataset, the 3DSSD takes about 100ms per scene because of more input points.

I haven't tested the codebase on NuScenes dataset because I don't have enough GPUs right now to train the model which needs 4 Nvidia V100 with 32G GPU memory for...

Hi, This is not our original codebase, and I haven't tested the part of multi-class training. Since I am busy on some other tasks recently, I don't have time to...

Hi @benjaminrwilson , thanks for the help! I align the camera and lidar timestamp according to the "_build_synchronization_cache" (av2.datasets.sensor.sensor_dataloader._build_synchronization_cache). * However, The timestamps of LiDAR and Camera are still not...

Hi, Thanks for your interests. I have no idea on that :(, since I haven't tried write CUDA with half precision support. Maybe check some existing code from other codebase,...

Hi, thanks for your interest. - Number of parameters: we have provided a trained EQ-Net on ScanNet. You can simply obtain the number of parameters through that pre-trained weight. -...

Yes. You can train one model on multi-task simultaneously with accordingly set query positions.

In my opinion, a query represents a position. You can use features of this position for detection, or for segmentation, or for both simultaneously, which depend on the head. If...

Seems like something is wrong with the chamferdist package. Maybe try installing the chamferdist package following the [4docc](https://github.com/tarashakhurana/4d-occ-forecasting)?