请问合并kitti2012和kitti2015的训练，是属于main还是fintine

[gyes00205]

Hi @1434205229 , I didn't implement dataloader for merged kitti12/15, you can use finetune.py to finetune model on kitti12 and kitti15 separately.

[1434205229]

Thank you for your reply, but I noticed that your experimental section included merging datasets and training.

[1434205229]

[gyes00205]

Hi @1434205229 , I am not the original author, I just re-implement this paper. You can find the original repository by the following link https://github.com/disco14/MDCNet In this implementation, I directly borrow the dataloader from PSMNet so I didn't implement merged dataloader. Most of the recent papers will finetune merged dataset and upload their results to KITTI server to get better performance.

[gyes00205]

Because my master thesis is related to stereo matching, I sometimes re-implement some papers myself as an exercise. If you are interested in this field, we can discuss it together. Thanks

[1434205229]

I see. Your project is great, I just started working on stereo matching. Can you leave a contact information for more communication in the future

[1434205229]

I have roughly looked at your code and the original author's code. May I ask if your 3D aggregation still uses PSMnet, and will the effect be better than the source code? Also, why do you want to change the application's finalpass dataset to a cleanpass dataset

[gyes00205]

I directly use the 3D aggregation from PSMNet, because the original author said that their 3D stacked hourglass network is similar to PSMNet, GwcNet and CSPN. As for the performance, my implementation hasn't surpassed the original author's results in the paper but I think different 3D aggregation architectures should not make too much difference in performance. Maybe there are some training strategies that I didn't considered. The reason why I choose cleanpass dataset is because PSMNet use cleanpass to train their model, I'm too lazy to re-download the finalpass dataset. But I later found that some papers said that using finalpass would be better than cleanpass because finalpass will have some blurry pictures, which is a kind of data augmentation.

The following is my contact information: Linkedin: https://www.linkedin.com/in/chialunhsu-67018221b/ email: gyes00205@gmail.com

[disco14]

属于微调

[disco14]

现在sceneflow上进行训练，之后在kitti上进行微调，我是在CSN方法基础上改的，结果没有问题。