Code of NYUD2

[Mannix-D]

Is the code of noise experiments on NYUD2 the same as text-image datasets?
I conducted the noise experiments on the NYUD2 dataset; however, the accuracy is lower than the paper's.

[QingyangZhang]

Hi, we randomly and independently add noise to RGB and Depth modality for NYUD2 experiments. That is, we apply torchvision.transforms.RandomApply(add_noise, p＝0.5) for each modality individually. Thus we should expect 25% samples with both modalities corrupted, 25% samples with RGB modality corrupted only, 25% samples with Depth modality corrupted only and 25% clean samples. Due to my current busy schedule, I will update the test code to GitHub later. Sorry for any inconvenience.

[Mannix-D]

. Thus we should expect 25% samples with both modalities corrupted, 25% samples with RGB modality corrupted only, 25% samples with Depth modality corrupted only and 25% clean samples.

Thank you for your response. I would greatly appreciate it if you could share your code on GitHub.

[QingyangZhang]

. Thus we should expect 25% samples with both modalities corrupted, 25% samples with RGB modality corrupted only, 25% samples with Depth modality corrupted only and 25% clean samples.

Thank you for your response. I would greatly appreciate it if you could share your code on GitHub.

Hi, I have updated the repo and add test.py in the RGBD-scene-recognition folder. The results are the same with I reported in the manuscript. Free feel to let me informed whether you still have troubles in reproducing. Thus I can continue help you to debug or close this issue.

[Mannix-D]

. Thus we should expect 25% samples with both modalities corrupted, 25% samples with RGB modality corrupted only, 25% samples with Depth modality corrupted only and 25% clean samples.

Thank you for your response. I would greatly appreciate it if you could share your code on GitHub.

Hi, I have updated the repo and add test.py in the RGBD-scene-recognition folder. The results are the same with I reported in the manuscript. Free feel to let me informed whether you still have troubles in reproducing. Thus I can continue help you to debug or close this issue.

Thank you for your generous sharing. But I noticed your adding noise code is same as text-image datasets. The package named the additional_transform.AddGaussianNoise is the same as the dataset.AddGaussianNoise, right? I conduct your experiment on this nyud2 dataset from scratch, but can't reach your accuracy.

[QingyangZhang]

That sounds weired. I also downloaded the datasets and trained from scratch myself. Here is my results:

• $\epsilon=0.0$, acc=70.33
• $\epsilon=5.0$, acc=61.47
• $\epsilon=10.0$, acc=54.89

Due to time limits, I only trained and tested once with random seed=1. The result is slightly different with that in the manuscript. Would you mind to share your code and accuracy?

As for the corruption function: Yes, AddGaussianNoise is the same for two tasks.

[Mannix-D]

That sounds weired. I also downloaded the datasets and trained from scratch myself. Here is my results:

• ϵ

  0.0
  , acc=70.33

• ϵ

  5.0
  , acc=61.47

• ϵ

  10.0
  , acc=54.89

Due to time limits, I only trained and tested once with random seed=1. The result is slightly different with that in the manuscript. Would you mind to share your code and accuracy?

As for the corruption function: Yes, AddGaussianNoise is the same for two tasks.

hello, I just notice that you are using the data.aligned_conc_dataset_noised import AlignedConcDataset as current dataset format. I will try the experiment again. Thanks!