Hi Thank you so much for the code release.

Can you please give the exact training and evaluation commands used for training and testing the ScanObjectNN dataset to recreate the results of table 2 in paper?

Thanks in advance. Much appreciated.

Hellow，Thanks for your nice work，there is a bug：

i did not change any parameters ，and run ：

CUDA_VISIBLE_DEVICES='2' python run_mvtn.py \
        --data_dir data/ModelNet40/ \
        --run_mode train \
        --mvnetwork mvcnn \
        --nb_views 8 \
        --views_config learned_spherical

output：

...
Evaluation:
        train acc: 82.79 - train Loss: 0.6117
        Val Acc: 76.54 - val Loss: 0.8504
        Current best val acc: 76.90

-----------------------------------
Epoch: [45/100]
        Iter [50/492] Loss: 0.4384
Remedy did not work

In paper,MVTN using resnet18 can get the accuracy of 90% on ModelNet40 test. I can only get 78% accuracy.My command is python run_mvtn.py --data_dir data/ModelNet40/ --run_mode train --mvnetwork mvcnn --nb_views 8 --views_config learned_spherical

Thank you for your last reply, I'm sorry I have another question.

Because it is not stated in the paper, I think that the ModelNet40 dataset used is unaligned.

However, in the supplementary material, it states that when testing rotation robustness,

A common practice in the literature in 3D shape classification is to test the robustness of models trained on the aligned dataset by injecting perturbations during test time [ 20 ]. We follow the same setup as [ 20 ] by introducing random rotations during test time around the Y-axis (gravity-axis).

Does this mean that you are using an aligned ModelNet40 dataset only for testing rotation robustness? Or is the ModetNet40 you stated in README.md already an aligned dataset?

Can you please confirm if the classification results are using 2048 input points instead of 1024 points?

Not clear from the tables (mainly because results of previous models are based on 1024 input points).

Kindly share the results of classification with 1024 points or provide exact commands and configurations to recreate the results.

Thanks in advance

Hey! I am new to this kind of approach and wanted to apply it to a custom dataset of mine. How do I find the correct nb_points value for my dataset? Thanks!

Thank you very much for your reply. Sorry I have encountered another problem. What is the use of the weights under results/checkpoints/modelnet/? Could you provide the model weights for the best results?

Using the current code on Blender 2.8 yields the following error:

bpy.data.objects['Camera'].select = True
AttributeError: 'Object' object has no attribute 'select'

According to API changes in Blender 2.8:

In 2.7x, you could directly (de)select an Object from its select property. This has been removed in 2.8x, in favor of some get/set functions.

Therefore, we need to change this piece of code to allow for direct usage on Blender 2.8, which is now the default for installation.

You know, in China, we can not access files on google drive. Can you share other download links to ModelNet40 and ScanObjectNN? For example, Baidu NetDisk has similar function with google drive, which is also widely used in China and many other countries.

It would be appreciated if you can provide such links. Thanks.

I trained a MVTN model with 100 epochs with the following command, and stopped training after 57 epochs.

python run_mvtn.py --data_dir data/ModelNet40/ --run_mode train --mvnetwork mvcnn --epochs 100 --nb_views 1 --views_config learned_circular

And the output of the 57th epoch is like this,

Epoch: [57/100] Iter [50/492] Loss: 0.7633 Iter [100/492] Loss: 0.7892 Iter [150/492] Loss: 0.3939 Iter [200/492] Loss: 0.1820 Iter [250/492] Loss: 0.2282 Iter [300/492] Loss: 0.6939 Iter [350/492] Loss: 0.4468 Iter [400/492] Loss: 0.2383 Iter [450/492] Loss: 0.5454 Evaluation: train acc: 82.03 - train Loss: 0.6457 Val Acc: 71.31 - val Loss: 1.0960 Current best val acc: 72.61

When I load the trained model to continue training, although it started training from the 58th epoch correctly, the accuracies got lower,

Epoch: [58/100] Iter [50/492] Loss: 1.2060 Iter [100/492] Loss: 0.6699 Iter [150/492] Loss: 0.5014 Iter [200/492] Loss: 0.4189 Iter [250/492] Loss: 0.2721 Iter [300/492] Loss: 0.3099 Iter [350/492] Loss: 1.0518 Iter [400/492] Loss: 1.1512 Iter [450/492] Loss: 0.2506 Evaluation: train acc: 55.48 - train Loss: 1.6519 Val Acc: 60.13 - val Loss: 1.4470 Current best val acc: 72.61

I found that in ops.py line 260-264, only when is_learning_views = True , the trained MVTN model will be loaded,

if setup["is_learning_views"]:
        models_bag["mvtn"].load_state_dict(
            checkpoint['mvtn'])
        models_bag["mvtn_optimizer"].load_state_dict(
            checkpoint['mvtn_optimizer'])

and in line 55-56, is_learning_views in setup is initialized like this,

setup["is_learning_views"] = setup["views_config"] in ["learned_offset",
                                                       "learned_direct", "learned_spherical", "learned_random", "learned_transfer"]

should the learned_offset in line 55 be repalced by learned_circular? Becaues the choices of learned views_config must be learned_circular, learned_spherical, learned_direct, learned_random or learned_transfer.

I am sorry if the reason is not here. I would appreciate it if you could tell me the correct way. :) @ajhamdi