rehg-lab / lowshot-shapebias Goto Github PK

Hi, thanks for the great work and for sharing your code!

I am just wondering, if possible, and convenient for you, could you also release the code for generating 2D-images/points-clouds from 3D models? Or provide the example code for doing these things (for example, the code snapshot for one single 3D model with given camera view parameters)? I am trying to reproduce your dataset building pipeline but I am not sure whether I do it in the same way as your paper. So I think a code example would be really helpful for us to follow your work!

Thanks for the insightful paper and great dataset again!

Hi, thanks for the great code and for sharing your dataset. It seems a file is missing. Can you please look into it?

File "github/lowshot-shapebias/lssb/nets/__init__.py", line 4, in <module>
    from .conv4 import Conv4 as conv4
ModuleNotFoundError: No module named 'lssb.nets.conv4'

Did anyone meet this issue and know how to solve it? Thank you in advance!

Traceback (most recent call last):
File "lssb/lowshot/train.py", line 112, in
main()
File "lssb/lowshot/train.py", line 61, in main
model = getattr(lowshot_models, hparams.model_type)(hparams)
File "./lssb/lowshot/models/image_simpleshot_classifier.py", line 7, in init
super().init(hparams)
File "./lssb/lowshot/models/base_simpleshot_classifier.py", line 19, in init
self.hparams = hparams
File "/home/myang47/anaconda3/envs/ifsl_pytorch1.7.0/lib/python3.8/site-packages/torch/nn/modules/module.py", line 826, in setattr
object.setattr(self, name, value)
AttributeError: can't set attribute

Hi guys,

Thanks for sharing this interesting work.
May I know more detailed operations about generating multi-view images from 3D objects in Blender? Or could you offer the scripts of this generation procedure kindly? My e-mail is [email protected].

Thanks again!

Hi @sstojanov ,

May I know how to pre-train DGCNN? It is hard for me to find the correct .sh and .py files to get DGCNN pretrained. Besides, may I know whether you pre-train DGCNN on all training classes in a traditional way or in a few-shot meta-learning way? Thanks in advance.

Minmin

Thank you for your work and for sharing the implementation.
But there are some issues, so please share your insight!

(1) Protocol of the whole experiment
Actually, it isn't easy for me to catch up on the proper reproduction protocol.
Please check whether I understand correctly or not.

1. Train SimpleShot point cloud classifier only.
   (bash training_scripts/train_dgcnn_simpleshot.sh 0)
2. Extract point cloud embeddings using the trained one.
   (python lssb/feat_extract/extract_simpleshot_pc_feature_modelnet.py --ckpt_path=[])
3. Train SimpleShot image classifier using the extracted embeddings.
   (bash training_scripts/train_resnet18_joint_simpleshot.sh 0)
   Here, the original code uses the config "modelnet-joint-simpleshot-resnet18-cfg", but it makes error. I think this should be "modelnet-joint-simpleshot-resnet18-w-pc-cfg" (or "modelnet-joint-simpleshot-resnet18-wo-pc-cfg"). Is it right?
4. Test trained image classifier.
   (bash testing_scripts/test_simpleshot_modelnet.sh)

(2) Performance
The number in the table in ReadMe is different from the table in the main paper.
For example, in the case of ModelNet 1shot-5way, LSSB achieves 61.91 in the main paper.
But in ReadMe, it is written as 57.57, which is much lower than the baseline (SimpleShot, 58.99).
What makes this huge difference?

Further, when I tested the officially provided checkpoints, it returns even lower performance (54~55).
I run five times as the paper recommended.
For this, I used "bash testing_scripts/test_joint_simpleshot_modelnet.sh".
Here, I slightly changed the sh file as follows, since there is no pretrained_models/simpleshot/modelnet/shape-biased directory.
python lssb/lowshot/test.py --log_dir=pretrained_models/simpleshot/modelnet/shape-biased-w-pc/
--name=joint-modelnet-pairwise-simpleshot-w-pc
--version=0
--gpu=1 \

It would be greatly appreciated if you could help me with these issues.