Question about training process and evaluation

[lolrudy]

I ran the training codes for all 6 objects in NOCS dataset and trained a detection model using CAMERA dataset. Excluding laptop, other objects does not work well using evaluation code from Shape Prior Deformation. I'm wondering whether hyperparameter or training process should be changed for other objects. Moreover, can you please release your evaluation code and trained model weight?

[codewfun]

Same problem, we trained the model with preprocessed data following Shape Prior Deformation and evaluate the trained model with evaluation code from that repo but only get the following results, which are far from the paper results. I'm wondering if there are other tricks that are missing apart from the released designed model and shape deformation data aug.

	IOU 50	IOU 75	5 deg, 5cm	10 deg, 5cm
paper result	92.2	63.5	28.2	64.6
Our implement	47.7	33.3	19.3	41.8

[codewfun]

@lolrudy Did you get better results and can you share your result?

[lolrudy]

I did not get anything meaningful... The result is even worse than yours. I doubt that there are bugs in the releasing code.

[codewfun]

@DC1991 Could you provide some advice? Thanks.

[makangzhe]

When i run gen_pts.py after file obj convert to file ply, i can't generate labeled data , can you help me ? can you share how do you train this model ? @codewfun

[codewfun]

I didn't use the gen_pts.py and the ply objects. Instead, I use the label and sampled model points from the Shape Prior Deformation repo. I think the pre-processing won't affect the performance much. The performance of our trained model are far from the paper results. I'm waiting for the author's reply.

[lolrudy]

We reimplement FS-Net and find out it actually works. The result is similar to the paper.

[taeyeopl]

@lolrudy
Q1. May I ask what part is different from the first version (worse performance)??
Q2. Can you share your reproducing code? include training & evaluation code & pretrained weights??
It would be really helpful.

[lolrudy]

Sorry I can't share the code in this moment. We preserve the network architecture, and rewrite the training code. The evaluation code is from shape prior deformation. There might be problem in the computation of rotation matrix from two vectors in the original code, but I'm not sure. Also, it should be careful to recover the translation and size from the network output.

[codewfun]

@lolrudy What do you mean "There might be problem in the computation of rotation matrix from two vectors in the original code"? Could you share more detail on the difference about this part?

[codewfun]

@lolrudy I tried fix the rotation matrix part and get better results. But there are still some gap to the paper result. Could you share your result? And did you use the detection result from shape prior deformation or YOLO v3 during evaluation?

[dedoogong]

@DC1991 , @lolrudy I also failed to reproduce the paper performance
My train results are as below(I didn't use CAMERA, but used only REAL dataset)

	mAP
3D IoU at 25:	79.32
3D IoU at 50:	64.38
3D IoU at 75:	17.48
5 degree, 2cm:	0.04
5 degree, 5cm:	0.14
10 degree, 2cm:	1.36
10 degree, 5cm:	3.76

Please help me find where is the buggy point in the training code~!

[amaj17]

@dedoogong can you share your evaluation code with me? I'm having troubles with evaluating the results using the code from the Shape Prior Deformation repository and it seems like you got it to work...