Pro-NeXt: An All-in-One Unified Model for General Fine-Grained Visual Recognition

Area of Science:

• Computer Vision
• Machine Learning

Background:

• General visual classification (CLS) struggles with specialized image recognition.
• Fine-Grained Visual Classification (FGVC) addresses this but existing methods lack generalizability and require task-specific models.
• Current FGVC benchmarks are often limited to homogeneous datasets.

Purpose of the Study:

• To propose a scalable and explainable foundational model for diverse FGVC tasks.
• To develop a unified approach that overcomes the limitations of task-specific models.
• To demonstrate the generalizability of the proposed model across disparate professional fields.

Main Methods:

• Introduction of a novel architecture named Pro-NeXt.
• Evaluation of Pro-NeXt on 12 distinct datasets across 5 diverse domains (fashion, medicine, art, etc.).
• Analysis of Pro-NeXt's scalability and explainability through intermediate feature performance.

Main Results:

• Pro-NeXt-B (basic size) surpasses all prior task-specific models on 12 datasets.
• Pro-NeXt exhibits strong generalizability across diverse professional fields.
• Scaling Pro-NeXt consistently improves accuracy, demonstrating good scaling properties.
• Intermediate features provide reliable object detection and segmentation without additional training.

Conclusions:

• Pro-NeXt offers a scalable, explainable, and generalizable solution for FGVC.
• The model's performance and adaptability across varied domains represent a significant advancement.
• Pro-NeXt's explainability through feature analysis opens new avenues for research.