problem about retraining the model

[zhzxlcc]

Excuse me, I want to retrain the model in the Flyingthings3D dataset and have set the parameters according to the paper. But I find a strange thing: step=85000 and 475000, the optical flow epe error is reduced from 12 to 7. The disparity and disparity change epe error are still at 42 and didn't fall down. What's the problem behind it?
Besides, I wonder to know how to use the knowledge distillation for better performance.
Thanks very much.

[FilippoAleotti]

Hi,
The network requires a long training time (especially in the synthetic domain). However, what do you mean by 42% epe error? If the network has a 42 epe error, there is something wrong (maybe in data or in training setup).
About distillation, I used it only in the real domain

[zhzxlcc]

Thanks for you reply.
Yes, the network has a 42 epe error with respect to the disparity and disparity change. Since ./configurations/sceneflow.json lacks the keys and parameters for training mode, I have set up the configuration on flyingthings according to the paper. Maybe something was missing. Could you please provide the training setup for sceneflow.json? Thanks a lot!

[zhzxlcc]

Sorry to bother you again~
With the help of you, I have almost achieved the same performance on Flyingthings3D. However, when trying to finetune the model on KITTI just with the provided gt, I got the results as follows:F1-27%,D1-6%,D2-9%(Noc);F1-36%,D1-6%,D2-15%(All).
I think the flow outliers are far from the error 18.53% in the paper obviously. Is the ablation study of KITTI in terms of Noc or All? And I wonder to know if you employed cross-validation or prolong the training process or just finetune the model with the kitti.json? Some suggestions are helpeful~
Thanks very much!

[FilippoAleotti]

Hi,
The results are about OCC validation of the model fine-tuned on KITTI

[zhzxlcc]

Thanks a lot! I still have problems about the performance on KITTI dataset. The error is far from the normal value. I'll check it. Thanks for your attention to this matter.

[FilippoAleotti]

Hi,
Are you using kitti_160 split right? And NN as interpolation method in filter_settings?

[zhzxlcc]

In fact, I create txt for kitti_170 split according to your kitti_160.txt and use the last 30 pictures as validation. And do you mean I should modify "NN as interpolation" in "filters_setting{method :NN}" of the kitti.json in the configurations?
I didn't modify the parameters in the kitti.json before. After setting the filters_setting:method as NN, I got the following results with respect to OCC: F1-33%,D1-5%,D2-13%. Maybe something wrong with corr2d when run "python ops.py“. It is strange that only flow outliers is abnormal.
Thanks a lot !

[FilippoAleotti]

Are you using the evaluate_partial_scene_flow script inside external_packages/kitti_test_suite for evaluation (make sure to change it if you are using 170 images for training, and not 160)?

[zhzxlcc]

Thanks for your attention. For valuation, I use the KITTI official code implemented in MATLAB. By the time, I got the OCC results for 30 valuation set as follows: F1-22%,D1-4%,D2-9%(All). I think they are closer to the results only using KITTI gt in your awesome paper~ Would you consider it is the reasonable result?
I am very grateful for your guidance.