The paper "Part-based R-CNNs for Fine-grained Category Detection" by Ning Zhang, Jeff Donahue, Ross Girshick, and Trevor Darrell introduces a novel method for fine-grained category detection that leverages deep convolutional features computed on bottom-up region proposals. The authors address the challenge of accurately localizing object parts, which is crucial for fine-grained recognition, by training both whole-object and part detectors and enforcing geometric constraints between them. Their method does not require bounding box annotations at test time, making it more robust and practical compared to existing methods. Experiments on the Caltech-UCSD bird dataset demonstrate that their approach outperforms state-of-the-art methods in fine-grained categorization tasks, even without ground truth bounding boxes. The paper also explores different geometric constraints and evaluates the effectiveness of part localization using selective search proposals. The results show that the proposed method achieves high recall for part localization, even in the fully automatic setting where the ground truth bounding box is unknown.The paper "Part-based R-CNNs for Fine-grained Category Detection" by Ning Zhang, Jeff Donahue, Ross Girshick, and Trevor Darrell introduces a novel method for fine-grained category detection that leverages deep convolutional features computed on bottom-up region proposals. The authors address the challenge of accurately localizing object parts, which is crucial for fine-grained recognition, by training both whole-object and part detectors and enforcing geometric constraints between them. Their method does not require bounding box annotations at test time, making it more robust and practical compared to existing methods. Experiments on the Caltech-UCSD bird dataset demonstrate that their approach outperforms state-of-the-art methods in fine-grained categorization tasks, even without ground truth bounding boxes. The paper also explores different geometric constraints and evaluates the effectiveness of part localization using selective search proposals. The results show that the proposed method achieves high recall for part localization, even in the fully automatic setting where the ground truth bounding box is unknown.