Model download link invalid

[Loucy-28]

Hi, thank you for your great work. I'm trying to reproduce the BinsFormer (Swin-L on NYU) results, but the Google Drive link for the pre-trained model seems to be broken. Could you please provide an updated link? Thanks!

[zhyever]

Sorry guys, I deleted the file on my google drive by mistake and cannot recover it now... I think I can re-train the model at the end of July and release it in August. Thanks a lot for your interest to this repo and my work. (this issue is also mentioned in #102)

[Loucy-28]

Mistakes happen. Thanks for keeping us updated—looking forward to the release in August.

[zhyever]

I found local backups for binsformer. Please check:

swinl: https://drive.google.com/file/d/1sZFsdU-q8QTmfMXeL5yLTHSPKl8ls6H6/view?usp=sharing

swint: https://drive.google.com/file/d/1bHgOnNcNLlQyyzFiBZao280cMjz7mhMy/view?usp=sharing

[Loucy-28]

Thanks for sharing the pre-trained models! I’m trying to train Binsformer on Gated2Depth data(using 3 stacked gated images instead of RGB). The loss decreases, but evaluation metrics stay poor (see attached log 20250721_182441.log).
Tried:
Using your pre-trained model
Switching to L1 loss
How should I modify it?

[zhyever]

Hi, after checking the log, I don't think the model is converging.

Please first check if the max_depth is correctly set to the max depth of Gated2Depth dataset.

Visualization of results might be necessary.

If it's still not converging, please try to re-run the code multiple times with different seeds. Sometimes, it can work.

[Loucy-28]

max_depth is correct, how to set different seeds?

[zhyever]

Please check:

Monocular-Depth-Estimation-Toolbox/tools/train.py

Line 44 in d8aff50

parser.add_argument('--seed', type=int, default=None, help='random seed')

Simply adding --seed xxx can work.

Given your loss is around 2.8, I guess a good run will reduce the loss to around 1.0 quickly (in the first several steps). If the loss is still around 2.8, simply kill and re-run.

[Loucy-28]

Okey,I'll try.

[Loucy-28]

Hi,the previous issue with the model not converging has been resolved, but I've run into a new problem: how can I visualize the color depth map output by the model? I used the commands in your inference.md file:

python tools/test.py configs/binsformer/binsformer_swinl_22k_w7_kitti.py
work_dirs/binsformer_swinl_22k_w7_kitti/20250728_133435/iter_38400.pth
--show

But the output isn't what I expected.(

) Did I make a mistake somewhere?How should I output the color depth map?

[Loucy-28]

Already modified