This an example of training the Marigold Depth Estimation using accelerator using the sceneflow dataset. Since the original training code is not open source, only the inference pipeline is released, so the performance is not guaranteed. BTW, Any other dataset is fine, just change the dataloader.
Reference Code: Marigold-ETH
Reference Paper: Marigold: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation
cd scripts
sh inference.sh
cd scripts
sh train.sh
Note the training at least takes 21 VRAM even the batch size is set to 1.