Explicit Correspondence Matching for Generalizable Neural Radiance Fields.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Neural Radiance Fields (NeRF) enable novel view synthesis but often require many source views.
• Generalizing NeRF to unseen scenarios and reducing view dependency remains a challenge.

Purpose of the Study:

• To develop a generalizable NeRF method capable of novel view synthesis with few source views.
• To improve geometry priors for NeRF by explicitly modeling feature correspondences across views.

Main Methods:

• Introduced a method utilizing explicit correspondence matching via cosine similarity of image features.
• Employed Transformer cross-attention to model cross-view feature interactions, enhancing matching quality.
• Integrated learned feature similarity as a geometry prior for NeRF color and density prediction.

Main Results:

• Achieved state-of-the-art performance on various evaluation settings.
• Demonstrated strong correlation between learned cosine feature similarity and volume density.
• Showcased the method's ability to generalize to new, unseen scenarios with as few as two source views.

Conclusions:

• The proposed method effectively uses explicit correspondence matching and cross-view attention for superior NeRF-based novel view synthesis.
• The learned feature similarity provides reliable geometric cues, improving NeRF performance and generalizability.
• This approach significantly advances the efficiency and applicability of NeRF for 3D scene reconstruction and rendering.