Prediction (initial depth and normal) seems to be better than estimation (final depth and normal)

[RexSan0x]

Hey @xjqi , thanks for sharing the code. When I ran the code.py, I have updated both the prediction files (depth_pred.mat and norm_pred.mat) and estimation files(depth_estimate.mat and norm_estimate.mat) for the 654 test images provided, to see the comparison in the evaluation between the predicted output (initial depth and normal) and estimated output (final depth and normal). However, after running the evaluation, I have found that the accuracy for prediction (around 98.9%) is more than the estimation (around 98.3%). I ran the test for the test_images for another 4 iteration and, even though the accuracy for prediction remains the same, the estimation is still lesser (around 98.3%). When i tried to visualise the depth and normal maps, the estimation seems to be more distorted than the depth prediction. Please suggest a solution for this problem or any step I have done wrong.

[xjqi]

Thank you for letting us know. As this has been a long time, I may need more time to diagnose this problem. I think the problem might be caused by the data loader.

[xjqi]

May I know what version of Tensorflow you use? This may also impact the results.

[RexSan0x]

Hi @xjqi, thanks for replying!! The Tensorflow version I have used is "1.13.2" and the python version is "3.6". When I tried to use a higher version, like "1.15.2", there seems to be a configuration problem with "tensorflow.contrib". I also tried to change the code so that Tensorflow ver2 works, like using "tf_slim" library for the "slim" commands, and using "tf.compat.v1" for all the Tensorflow version 1 commands, and changing the "tf.contrib.layers.xavier_initializer(uniform=False)" to "tf1.keras.initializers.GlorotNormal()". However, I'm not able to use the GPU with this version, as I think I might be using a CUDA version lesser than 11. Once again, I appreciate the time and effort you have taken to solve my query and I hope to hear your reply soon. Thankyou

[shicaiji857]

Hi,@RexSan0x,I encountered the same problem as you now. Have you managed to solve it now?

[xjqi]

Dear all,

We will check the results later. Sorry for the inconvenience.

Thanks

On Wed, Mar 6, 2024 at 3:48 PM shicaiji857 @.***> wrote:

@.*** https://github.com/RexSan0x,I encountered the same problem as you now. Have you managed to solve it now?

— Reply to this email directly, view it on GitHub https://github.com/xjqi/GeoNet/issues/8#issuecomment-1980269599, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACP5RTUDDIBRSRDWYFZWCFTYW3C4RAVCNFSM54H5Y7IKU5DIOJSWCZC7NNSXTN2JONZXKZKDN5WW2ZLOOQ5TCOJYGAZDMOJVHE4Q . You are receiving this because you were mentioned.Message ID: @.***>

-- Xiaojuan Qi, PhD Assistant Professor Department of Electrical and Electronic Engineering The University of Hong Kong

[shicaiji857]

Hi@xjqi，I would like to know if you have plans to release the 'GEONET++' code for us to study. I look forward to your reply.

[xjqi]

Hi@xjqi，I would like to know if you have plans to release the 'GEONET++' code for us to study. I look forward to your reply.

Geonet ++ has already been incorporated in the code. See edge refinement