Questions about ViewFormer Implementation: bin File Generation, Visualization Methods, and Single GPU Training Configuration

[DDEONSIK]

Questions about ViewFormer Implementation: bin File Generation, Visualization Methods, and Single GPU Training Configuration

Hello, I have a few questions regarding ViewFormer:

1. Bin File Generation for Oviz Visualization

I understand that a bin file format is required to visualize training results using Oviz. Could you explain how to generate these bin files?

2. Visualization Methods

How did you create visualization images like Figure 5 in the paper? If it was generated using code, would it be possible to share the demo code?

3. Single GPU Training Configuration

In the viewformer_r50_704x256_seq_90e.py file, I noticed that changing the num_gpus value affects the total number of iterations during training. When training with a single GPU, should I set num_gpus to 1 and increase the number of iterations to achieve performance levels comparable to those reported in the paper?

(Note: Using the default value of num_gpus=8 did not achieve the performance levels mentioned in the paper.)

Thank you for your help!

[ffengc]

I have been following this excellent project recently and noticed that you are also interested in its visualization methods. I would like to ask you @DDEONSIK—and the author @ViewFormerOcc—how exactly the results of this project are intended to be visualized. I didn’t find any detailed instructions in the README file. After calling the Evaluation method, I only see numerical results and not the actual point clouds generated for the test set. Could you please explain how to visualize the outputs of this work?
Thank you very much for your help! 🌹

[DDEONSIK]

I have been following this excellent project recently and noticed that you are also interested in its visualization methods. I would like to ask you @DDEONSIK—and the author @ViewFormerOcc—how exactly the results of this project are intended to be visualized. I didn’t find any detailed instructions in the README file. After calling the Evaluation method, I only see numerical results and not the actual point clouds generated for the test set. Could you please explain how to visualize the outputs of this work? Thank you very much for your help! 🌹

About 1. Bin File Generation for Oviz Visualization.

Summary: I also need help with this issue.

Hello, I understand that your situation is very similar to mine regarding the Oviz visualization process.

Like you, after model training, validation, and testing, I only have the latest.pth and training .log files, with no additional information. While there is code available to generate .pkl files from latest.pth, I haven't attempted it as it doesn't seem helpful for visualization purposes.

I'm currently working on writing new code to convert our limited information into Occ data .bin / .pcd / .txt that can be input into Oviz, but I'm not making much progress.

Results of running the Oviz Demo:

Like the visualization results in the paper Fig.6, I'm aiming to generate .bin files that would produce the most similar visualization, but I'm experiencing difficulties. According to the Oviz README, the occ data should be in N*6 format x, y, z, cls, vx, vy, float32, default size 0.4m.

I wish I could help more, but unfortunately I'm in the same situation. If you find a solution, please share it. Thank you.

[ffengc]

Thank you for sharing! @DDEONSIK

[ViewFormerOcc]

see #10