point sampling problems about prnet HOT 7 CLOSED

hi @lx709 Did you also train the model with random_p2=random_p1, if yes, then how is the performance? I'm also checking the code and it seems there are two settings to simulate the "partial" scan.

from prnet.

@ShengyuH Please check my answer above, when random_p2=random_p1, the performance is:
A->B:: Stage: best_test, Epoch: 58, Loss: 0.018334, Feature_alignment_loss: 0.001463, Cycle_consistency_loss: 0.002618, Scale_consensus_loss: 0.000000, Rot_MSE: 9.122054, Rot_RMSE: 3.020274, Rot_MAE: 1.405421, Rot_R2: 0.945782, Trans_MSE: 0.000274, Trans_RMSE: 0.016538, Trans_MAE: 0.010954, Trans_R2: 0.996694.

from prnet.

Did you train using the default parameters? Or could you please share your set of parameters for training?

from prnet.

yes, i'm using default parameters.

from prnet.

@lx709 I just reproduce the same behavior when I train PRNet on another dataset.

1. random_p2=random_p1:
   A->B:: Stage: best_test, Epoch: 73, Loss: 0.050626, Feature_alignment_loss: 0.006549, Cycle_consistency_loss: 0.002129, Scale_consensus_loss: 0.000000, Rot_MSE: 25.722961, Rot_RMSE: 5.071781, Rot_MAE: 3.447572, Rot_R2: 0.844065, Trans_MSE: 0.002287, Trans_RMSE: 0.047822, Trans_MAE: 0.034319, Trans_R2: 0.972008

2. random_p2!=random_p1
   A->B:: Stage: best_test, Epoch: 99, Loss: 0.105203, Feature_alignment_loss: 0.013792, Cycle_consistency_loss: 0.004159, Scale_consensus_loss: 0.000000, Rot_MSE: 72.986801, Rot_RMSE: 8.543231, Rot_MAE: 5.605420, Rot_R2: 0.590017, Trans_MSE: 0.007507, Trans_RMSE: 0.086643, Trans_MAE: 0.061235, Trans_R2: 0.909541

from prnet.

Hi, @WangYueFt Thanks for sharing your code! In your paper, you claim the method to handle partial-to-partial point set registration. But under your setting, especially I checked your dataloader, it seems you only sample one point, and get the nearest 768 points for both the source and target shape. In my opinion, this will generate two point sets with large overlap, especially you pick the point from a very large distance from the shape (random_p1 = np.random.random(size=(1, 3)) + np.array([[500, 500, 500]]) * np.random.choice([1, -1, 1, -1])). I visualized some pairs and found the source point sets and target points have a large part of overlaps.

When I generate source and target point sets with two different samplings, i.e., random_p2 is picked instead of 'random_p2 = random_p1', I got the following performance, which falls a lot behind the one you report in the paper.

exp1: identical sampling for source and target:
A->B:: Stage: best_test, Epoch: 58, Loss: 0.018334, Feature_alignment_loss: 0.001463, Cycle_consistency_loss: 0.002618, Scale_consensus_loss: 0.000000, Rot_MSE: 9.122054, Rot_RMSE: 3.020274, Rot_MAE: 1.405421, Rot_R2: 0.945782, Trans_MSE: 0.000274, Trans_RMSE: 0.016538, Trans_MAE: 0.010954, Trans_R2: 0.996694

exp2: independent point sampling for each shape
A->B:: Stage: best_test, Epoch: 94, Loss: 0.036311, Feature_alignment_loss: 0.003807, Cycle_consistency_loss: 0.002531, Scale_consensus_loss: 0.000000, Rot_MSE: 32.466419, Rot_RMSE: 5.697931, Rot_MAE: 3.089283, Rot_R2: 0.807019, Trans_MSE: 0.001563, Trans_RMSE: 0.039535, Trans_MAE: 0.026601, Trans_R2: 0.981015

Hi,

Let me try to answer this question. Yeah, we simulate the the partial-to-partial case by taking the closest points from both pointcloud1 and pointcloud2. If using the sampling commented out in the code, that will make the task harder and we expect worse results. Hope this helps.

Best,
Yue

from prnet.

got it, thanks for your kind reply @WangYueFt .

from prnet.