Code for Recovering Part Order

[absdnd]

Hi,

Thank you for releasing the code for your approach.

Could you also release the code for Part-Denoising (Fig 10 in the PQ-NET paper)?

[ChrisWu1997]

Hi Abhinav,

Part order denoising is a simple future application that we demonstrated. I sort of lost the code for that experiment now, but it only needs several lines of code.

Before the following line,

PQ-NET/agent/agent_seq2seq.py

Line 76 in 50e3a95

output_seq, output_stop = self.net(part_feature_seq, target_seq, self.teacher_forcing_ratio)

randomly shuffle part_feature_seq to create noisy part order. Note that each data in a single batch has different number of parts, so you need to may do this using a for loop. The model needs to be re-trained.

[absdnd]

Thanks for your reply.

Could you tell me if the part-denoising experiment would be accurate enough to be used in another application which needs canonical part orders?

[ChrisWu1997]

It depends on the data. Since it's trained in a supervised manner, I think it cannot generalize well if using it on other data distribution that is much different from PartNet.

[absdnd]

Thanks.

Just one more question. In the code segment above, the network would generate novel 3D shapes in the canonical part order. However, is it possible only to generate a permutation of the input parts that yields the denoised order?

[ChrisWu1997]

If I understand correctly, you are asking to generate only a permutation of the input parts, i.e. a sequence of index. Our model cannot do this, but I think you may look into this paper Pointer Network, which aims to solve a similar re-ordering problem.

[absdnd]

Sure, thanks. Is it possible to utilize your architecture and perform linear assignment on the generated shapes to learn the part order? I mean matching the generated shapes and the input shapes.

[ChrisWu1997]

I think it is possible, by combining our framework with Pointer Network. But I'm not sure how it works in practice.