kamse / costdcnet Goto Github PK

Greetings! I appreciate your remarkable work and would like to inquire about the experimental setup described in Table 2.

Specifically, I am curious about the environment in which you compared Point-Fusion and the method proposed in your paper when the number of points is 32.

We attempted to replicate your approach by using the provided CostDCNet model (available at https://github.com/kamse/CostDCNet/tree/main/weights) for inference, but we could not achieve the same level of performance reported in your paper. Consequently, we surmised that the results in the paper were obtained by training CostDCNet with 32 points instead of 500. Could you clarify whether this is the case?

Hi~ kamse,
Thank you for your excellent work! It's very inspiring to me. Could you provide your training code?

Hello, thanks for the great work! It inspires me a lot.

Could you provide the pretrained models for the three datasets?

Thank you for the interesting work! I just had a quick question regarding the Ablation study in Table 4.

Going from Row 1 to Row 2, the 2D (RGBD) UNet is replaced with the proposed 3D UNet. I was wondering - how does the # of parameters decrease when going from 2D to 3D UNet? I would have thought that going from 2D to 3D convolutions, the # of parameters would increase. More specifically, what is the structure of the 2D UNet? And, am I correct in assuming that the loss for the 2D UNet is simply direct regression + L1 instead of soft-argmax + L1?

Also, before the 2D & 3D encoders are added, when generating the fused volume, what exactly are the 2D and 3D features used?

none