Open Choi-YeongJoon opened 6 months ago
The depth anything metric distance model is different from the normal depth-anything model. Make sure you are downloading the correct version and finetuning the correct version. I suggest going through zoedepth training instructions.
Thank you for your advice! I followed your advice, finally, looked at the zoedepth guide, and succeeded in parsing onnx. I am then trying to parse with tensorrt. I'll let you know how it goes!
Tensorrt has some pitfalls including precision point, beware of how the ONNX model translates that for your pytorch model which is I am guessing will be trained at float32 or 64.
Thank you for your sensitive response. Depending on the requirements given to me, I think I can use a sufficiently large precision. However, there are environmental restrictions. It is a limitation of the Tensorrt version, and there appear to be some operations that are not supported in version TRT 8.2. I'm going to construct a custom operation and parse it. Do you have any experience doing this before?
No, I am afraid, I have no expertise with this. But you should check out the ONNX backend for Tensorrt using the execution provider.
The only unsupported operation is cubic interpolation. This is supported in version 8.6 or higher, but is not supported elsewhere. If it is resolved, I will share it.
That actually is good enough.
hi , can you tell me how to download the dataset and how to fine tune the depth anything to get metric depth estimation ? the instruction in this repository isnt clear for me :(
Hi, I'm very happy that you solved the problem. I also trained my dataset to obtain finetuned model files, but I don't know where to use the model files to predict images and obtain depth maps. Do you know how to solve this problem? Thank you.
I fine-tuned pretrained model(depth_anything_vits14.pth) using KITTI Data set, I could get finetuned model (depth_anything_finetune/ZoeDepthv1_05-Mar_14-45-9b59bd15407a_best.pt) And I confirmed that this infers the metric depth quite accurately! As a result, How can export this model to onnx file? I exported pretrained model(depth_anything_vits14.pth) before, but fine-tuned model, doesn't work.
Error message is below :
I expected the result of the fine-tuned model to be depth-anything, but Upon checking, it appears that the zoedepth model is being trained. Is it possible to make absolute value output from depth-anything itself? and Is there a script that can do inference using a fine-tuned model?
Anyone who has success with this please help me.