Coarse-to-Fine Multi-Scene Pose Regression With Transformers.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Absolute camera pose regression estimates camera position and orientation from images.
• Current methods typically train on single scenes or use fully connected layers for multi-scene learning.
• Transformers offer a novel approach for aggregating image features and scene embeddings.

Purpose of the Study:

• To develop a Transformer-based model for multi-scene absolute camera pose regression.
• To enhance localization accuracy by focusing on generalizable features.
• To enable parallel processing of multiple scenes for efficient pose estimation.

Main Methods:

• Utilizing Transformer encoders for self-attention-based aggregation of activation maps.
• Employing Transformer decoders to transform latent features and scene encodings into pose predictions.
• Introducing a mixed classification-regression architecture to refine localization accuracy.

Main Results:

• The proposed Transformer model effectively learns to embed multiple scenes in parallel.
• The mixed classification-regression approach significantly improves localization accuracy.
• The method outperforms existing multi-scene and state-of-the-art single-scene pose regressors on benchmark datasets.

Conclusions:

• Transformer-based architectures are well-suited for multi-scene absolute camera pose regression.
• The novel approach achieves superior performance compared to previous methods.
• This work advances the field of camera localization with improved accuracy and efficiency.