search

weiyithu / SurroundDepth

[CoRL 2022] SurroundDepth: Entangling Surrounding Views for Self-Supervised Multi-Camera Depth Estimation
MIT License
247 stars 38 forks source link

Reproduction problem of the nuscenes dataset #10

Open XiaofengLi-HKUST opened 1 year ago

XiaofengLi-HKUST commented 1 year ago

Hello, Your results are great, I'm following your work. I have a problem of reproducing the results for the nuScenes dataset, the abs_rel by Monodepth2 for the nuScenes dataset are high to 0.35, and the curve during training is very strange. I use the code of Monodepth2, and use your code to process the dataset. How can I reproduce the Monodepth2 results in your paper？Are there any details that need special attention? Thanks！ 1665560657308 103

weiyithu commented 1 year ago

Hi, in my experiment, I find the models are easy to be overfitted on nuScenes. However, high to 0.35 is still strange. Do you use my nuScenes dataset list? where the static frames are removed.

XiaofengLi-HKUST commented 1 year ago

yes, I use your nuScenes data list. Is there any difference in detail between training DDAD and nuScenes beside removing static frames and image resolution？

weiyithu commented 1 year ago

Except some hyperparameters, I think there does not exist differences. Maybe the only difference is that nuscenes is much easier to overfit.

XiaofengLi-HKUST commented 1 year ago

I have another question, I also can't reproduce the results of PackNet model. I find that the results of PackNet model in your paper is very different from the result of his original paper in DDAD dataset, may I ask whether the result of PackNet in your paper is reproduced by youself? Thanks!

weiyithu commented 1 year ago

Yeah, it is produced myself since the original paper only report front view results but we need all views' results.