Blake Chen

Hi, I hope you are using the same experimental setting to ours. It seems that your result is from the 3DSSD baseline. To the best of our knowledge, the performance...

Learning rate is a very important factor. If you are using a total batch size of 4, it is recommended to set the learning rate around `0.0025`.

Hi, Yes, we've made some modifications to the original 3DSSD. We noticed incorporating axis-aligned IoU loss and using cosine learning rate schedule (step schedule in the original version) could help...

Hi, It is feasible to train or evaluate the model with custom datasets, which is supported by the [OpenPCDet framework](https://github.com/open-mmlab/OpenPCDet). You can check [this document](https://github.com/open-mmlab/OpenPCDet/blob/master/docs/DEMO.md) to learn how to transform...

Hi, We didn't observe large performance difference between dividing the total batch size of 16 into 8x2 and 4x4. If you would like to use the 4x4 division, you could...

Hi, the pre-trained models provided in this repo are all trained under the commonly used train-val setting (3,712 samples for training and 3,769 samples for validation). We use the 80%-20%...

Hi, Yes, we randomly pick 80% labeled examples for training and use the remaining 20% for validation. You can simply write a script to generate the split or just follow...

Hi, the detailed implementation of `furthest_point_sampling_weights` is in [sampling_gpu.cu](https://github.com/blakechen97/SASA/blob/main/pcdet/ops/pointnet2/pointnet2_batch/src/sampling_gpu.cu). I haven't found the non-cuda implementation of the furthest point sampling algorithm and its variants. I'm afraid its non-cuda implementation is...