CATs++: Boosting Cost Aggregation With Convolutions and Transformers.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Cost aggregation is crucial for disambiguating noisy scores in image matching.
• Existing CNN-based methods struggle with severe deformations and limited receptive fields.

Purpose of the Study:

• To introduce Cost Aggregation with Transformers (CATs) for robust image matching.
• To propose CATs++, an efficient extension of CATs, to overcome computational limitations and enhance performance.

Main Methods:

• Developed CATs utilizing self-attention mechanisms for global context and appearance affinity modeling.
• Introduced CATs++ with early convolutions and a novel transformer architecture for efficient cost aggregation.
• Incorporated multi-level aggregation and swapping self-attention for improved accuracy and regularization.

Main Results:

• CATs demonstrated competitive performance but faced high computational costs.
• CATs++ achieved significant performance gains and reduced computational costs, enabling higher resolution processing.
• The proposed CATs++ method outperformed previous state-of-the-art methods by large margins.

Conclusions:

• CATs++ offers an efficient and effective solution for cost aggregation in image matching.
• The method's ability to handle high-resolution inputs and its superior performance make it a valuable advancement.
• The integration of transformers with convolutional inductive biases proves highly effective.