Code repo for the paper Semantic Correspondence via 2D-3D-2D Cycle.

Please run demo.py.

You can download them from Google Drive.

Training the full pipeline is somewhat involved and complicated, and our code is heavily based on ShapeHD. In general, there are four steps:

• Train ShapeHD model as outlined here.
• Prepare synthetic ShapeNet model renderings by mitsuba and generate their corresponding viewpoints through preprocess.py.
• Train viewpoint estimation network by running scripts/train_vp.sh.
• Train 3D embedding prediction network by running train_embs.py and then generate keypoints' average embeddings for retrieval. This step requires KeypointNet dataset.