[Uploaded model] Depth-Anything-Small metric depth model

[lukasz-staniszewski]

Hi, just letting you know I've trained Depth-Anything-Small for KITTI metric prediction (10 epochs).

Feel free to re-use/finetune it or put it in the repo:

Wandb run HuggingFace checkpoint

Metrics:

Method	$\delta_1 \uparrow$	$\delta_2 \uparrow$	$\delta_3 \uparrow$	AbsRel $\downarrow$	RMSE $\downarrow$	log10 $\downarrow$
ZoeDepth	0.971	0.996	0.999	0.054	2.281	0.082
Depth Anything Large	0.982	0.998	1.000	0.046	1.896	0.069
Depth Anything Small	0.971	0.996	0.999	0.053	2.267	0.023

Model size:

Method	model size (MB) $\downarrow$
Depth Anything Large	1 344.4
Depth Anything Small	101.6

[gopin95]

怎么推理

[daegyu93]

I want to use a model you've created, but in the new model you made, there are additional parameters like

"conv2.weight", "conv2.bias", "seed_bin_regressor._net.0.weight", "seed_bin_regressor._net.0.bias", "seed_bin_regressor._net.2.weight", ..... 

and so on. Because of this, I'm unsure how to properly load the model. When using DepthAnythingV2 and attempting to use load_state_dict, I encounter the following error:

RuntimeError: Error(s) in loading state_dict for DepthAnythingV2:
    Unexpected key(s) in state_dict: "conv2.weight", "conv2.bias" ......

How can I resolve this issue?

[lukasz-staniszewski]

For metric depth, we need to load the weights using the metric_depth code. You can follow this file to use the weights.

Here is how I use it in my project, where the process_image function returns numpy array of metric depths: LINK.

[ZuoJiaxing]

@lukasz-staniszewski Hi, thanks for sharing the trained model! Can you share the config_zoedetph.json file you used to train it? Thanks a lot!

[lukasz-staniszewski]

@lukasz-staniszewski Hi, thanks for sharing the trained model! Can you share the config_zoedetph.json file you used to train it? Thanks a lot!

I'm pretty sure I didn't change the file, and it's the same as this one. However, there is also config_zoedepth_kitti.json, and I'm not sure which one is the exact one used under the hood.