tjqansthd / gdn-pytorch Goto Github PK

Hello，

When I did the evaluation test, I had some problems. I used KITTI dataset segmented by eigen for evaluation, a total of 697 images. Since the code you provided contains some file names that belong to you, I wrote an independent evaluation code. But the results I got are not ideal, and they are very different from the results provided in your paper. I modified my code based on the compute_errors function in the calculate_error.py file in your code. Although the results have improved, there is still a gap with your results.

Based on the demo you provided, I wrote the prediction code to predict 697 images. I used all your operations, including up-sampling, scaling and channel transformation. I save the final result as an .npy file. The detailed process of my evaluation code is as follows:

1. Read the real depth map (the sparse depth map based on the radar scanner provided by KITTI) and the predicted depth map.
2. Scale the predicted depth map so that it has the same size as the corresponding real depth map( gt ).
3. Normalize the real depth map( gt ) and the predicted depth map and multiply both by 80.
4. Use the segmentation strategy provided by Garg et al. to cut the real depth map( gt ) and the predicted depth map. Here I am based on the size of the real depth map( gt ).
5. Modify the depth value in the predicted depth map that exceeds the threshold, the maximum value is 80 and the minimum value is 0.001.
6. Calculate median_scaling.
   For the third step, I refer to the code you provided.

In the end, the results I got are as follows, which is far from the results in your thesis.

***************** | abs_rel | sq_rel | rms | log_rms | a1 | a2 | a3
first-order result | 0.2913 | 2.1279 | 6.7395 | 0.6446 | 0.5039 | 0.7606 | 0.858
Ref your code | 0.2508 | 1.9023 | 6.6494 | 0.4613 | 0.5662 | 0.814 | 0.9053

Can you give me some suggestions on how I can operate to get the excellent results as provided in your paper.