MambaOVD: a Mamba-based open-vocabulary object detection method.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Open-vocabulary object detection (OVD) is crucial for autonomous driving and robotics.
• Transformer architectures are common in OVD but computationally intensive.
• Existing methods face challenges with efficiency and practical deployment.

Purpose of the Study:

• To propose MambaOVD, an efficient OVD method utilizing the Mamba architecture.
• To address the computational demands of transformer-based OVD models.
• To enhance the performance and applicability of OVD in real-world scenarios.

Main Methods:

• Developed MambaOVD with four modules: image encoder, text encoder, Mamba-based fusion, and detection head.
• Employed Mamba layers for efficient image-text fusion.
• Trained and evaluated on diverse datasets: Objects365, GoldG, LVIS minival, and AutoMine.

Main Results:

• MambaOVD demonstrated superior performance over state-of-the-art models like YOLO-World-S, GLIPv2_T, and DetCLIP_T.
• Achieved significant improvements in both qualitative and quantitative evaluations.
• Showcased the effectiveness of the Mamba architecture for OVD tasks.

Conclusions:

• MambaOVD offers a computationally efficient and high-performing alternative to transformer-based OVD methods.
• The Mamba architecture provides a promising direction for advancing OVD research.
• MambaOVD has strong potential for practical applications in robotics and autonomous driving.