Using Hydranet on Kitti dataset

[MarSpit]

Hi, I would like to use your pretrained Hydranet for depth prediction on the Kitti dataset, as you did in your paper. Unfortunately, in the repo I can only find the files for the NYU-D dataset. It would be great if you could also provide the code (e.g. dataloader and pretrained neural net) you used when you did the investigations regarding Kitti. Many thanks!

[adithyagaurav]

Hi,

Thanks for reaching out. In the repo I think you will find a file ExpKITTI_joint.ckpt https://github.com/adithyagaurav/Multi_Task_Learning/blob/master/ExpKITTI_joint.ckpt, these are the weights for this model trained on KITTI dataset, in fact if you clone the repo and execute "python3 inference_image.py", the script will automatically load the weights on the model and provide the output.

What you do need for a successful run however is the RGB Image data, I have tested on the images that you can find in attachments. I will add some debug code to the inference scripts which I think will improve the clarity of the code

Regards, Adithya data.zip https://drive.google.com/file/d/1PopR2kPhVTPkOD_8XFhNKx7hetfS0cNj/view?usp=drive_web

On Wed, 5 Jan 2022, 05:29 MarSpit, @.***> wrote:

Hi, I would like to use your pretrained Hydranet for depth prediction on the Kitti dataset, as you did in your paper. Unfortunately, in the repo I can only find the files for the NYU-D dataset. It would be great if you could also provide the code (e.g. dataloader and pretrained neural net) you used when you did the investigations regarding Kitti. Many thanks!

— Reply to this email directly, view it on GitHub https://github.com/adithyagaurav/Multi_Task_Learning/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AJ6L7K3SVXYSO4J2NOOE4DTUUQMQZANCNFSM5LJR3HHQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you are subscribed to this thread.Message ID: @.***>

[MarSpit]

Great, many thanks for the immediate help. I got it going on Kitti 👍 Best regards Marius

[adithyagaurav]

Glad to help

The training on Kitti Dataset is trickier though. You see there's a fundamental problem in Kitti, that there are only like 200 segmentation data points. While there are many many more data points for depth prediction.

This creates a big training problem when working with Kitti. The authors in the paper I have referenced used Knowledge Distillation to solve this problem you should check that, I am still looking in to how to achieve that however.

Thankfully they provided weights of network so I could test the model. So training part at this time is only available for NYU-D dataset.

Regards

On Thu, 6 Jan 2022, 10:08 MarSpit, @.***> wrote:

Great, many thanks for the immediate help. I got it going on Kitti 👍 Best regards Marius

— Reply to this email directly, view it on GitHub https://github.com/adithyagaurav/Multi_Task_Learning/issues/1#issuecomment-1006664635, or unsubscribe https://github.com/notifications/unsubscribe-auth/AJ6L7K3M22PTGHRC4FZWTVDUUWWAXANCNFSM5LJR3HHQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you commented.Message ID: @.***>