v2.0.0

We’re excited to announce ECoDepth v2.0.0, a major restructuring of our monocular depth estimation codebase. This release offers significant improvements to make training and inference more straightforward and flexible:

---

What’s New

• Integrated Model Downloading
  Automatically download and cache pretrained checkpoints—no more manual handling of files.

• Generic DepthDataset
  Load any custom dataset with ease, using a consistent API.

• PyTorch Lightning Integration
  Enjoy streamlined training, validation, and checkpointing, with support for ONNX and TorchScript exports.

• Config-Based Workflows
  JSON config files now replace bash scripts, improving clarity and maintainability.

• Simplified Dependencies
  Removed extraneous packages like mmcv. Installing ECoDepth is now much smoother.

• Single Model Download
  We no longer require separate downloads for Stable Diffusion, CLIP, and VIT checkpoints—one file is all you need.

• Three Separate Workflows
  Clearly divided into train/, test/, and infer/ directories for easier navigation.

---

Getting Started

1. Install PyTorch (with or without GPU support).
2. Install Dependencies

   pip install -r requirements.txt

3. Configure & Run
   • Training: python train/train.py --config train/train_config.json
   • Testing: python test/test.py --config test/test_config.json
   • Inference: python infer/infer_image.py --config infer/image_config.json

For more details, please see the updated README.

---

Note:

• The original code that generated the CVPR 2024 paper results is tagged as [v1.0.0].
• For new projects, we strongly recommend using v2.0.0 for its cleaner, more modular design.

Thank you for using ECoDepth! Feel free to open an issue if you have any questions or feedback.

v1.0.0: Original Codebase for CVPR Paper Results

This release contains the original codebase used to produce results for our CVPR paper, ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation. While this version retains the exact implementation used for our published work, we recognize that it is relatively complex for users.

To address this, we are currently preparing a new version with a greatly simplified workflow. The updated version will be released soon as v2.0.0 to make the codebase more user-friendly. Stay tuned!