I have used the following code to test my own dataset and obtain dynamic masks, but the results are not satisfactory:
# For your own dataset:
python dynamic_predictor/launch.py --mode=eval_pose_custom \
--pretrained=Kai422kx/das3r \
--dir_path=data/custom/images \
--output_dir=data/custom/output \
--use_pred_mask
Should I optimize the dynamic masks by fine-tuning certain parameters, or does the pre-trianed network already have the capability to predict dynamic masks?
Additionally, I am a bit confused about which parts of the network are frozen during your training. Based on my understanding, here’s how it works:
CroCo and MonST3R backbones are frozen.
The DPT head is optimized to obtain dynamic masks.
Could you please confirm if this is correct? I’d appreciate your guidance!
I have used the following code to test my own dataset and obtain dynamic masks, but the results are not satisfactory:
Should I optimize the dynamic masks by fine-tuning certain parameters, or does the pre-trianed network already have the capability to predict dynamic masks?
Additionally, I am a bit confused about which parts of the network are frozen during your training. Based on my understanding, here’s how it works:
CroCo and MonST3R backbones are frozen.
The DPT head is optimized to obtain dynamic masks.
Could you please confirm if this is correct? I’d appreciate your guidance!