Prerained model of SVR

[gxu-tz]

Thank you for your code.I have seen other issues that the pretrained model is without overlap loss and code of generator class in modelSVR.py corresponds to Phase 1.
So to produce better results for shape reconstruction,should I retrain the AE model on phase 0 → phase 3 and change the code of generator class in modelSVR.py and then retrain the SVR model?

[czq142857]

Yes. Phase 3 provides slightly better results. But to truly improve the result quality, you may want to sample more points on the shapes (in the data pre-processing step), because the current data with at most 16384 sampled points (inside+outside) per shape is not sufficient for good reconstructions.

[gxu-tz]

Thank you for the response.
1.Does sampling more points on the shapes in the data pre-processing step mean changing the code
batch_size_3 = 16 * 16 * 16 * 4 in 2_gather_256vox_16_32_64.py in IM-NET repository?
2.I have another question. BSP-NET can generate sharp edges, so can it only generate approximate polygons for circles, as shown in Figure 7 in the paper？

[czq142857]

1. Yes, and also batch_size_2.

2. Yes. One solution is to use quadric surfaces to replace planes, but it may lead to overfitting. See the following papers for references.
   (a) https://arxiv.org/abs/2106.14274
   (b) https://arxiv.org/abs/2104.05652
   (c) Code of (a): https://github.com/czq142857/BSP-NET-original/tree/master/bsp_fc_qs_deeper_2048q

[gxu-tz]

Thank you for your prompt reply,it helps me a lot!I will pay attention to other work of your team.