YUB-200: A Dataset for Fine-Grained Bird Recognition

The research on fine-grained recognition of bird images has received widespread attention, but currently there are few publicly available datasets of fine-grained bird images, and there are problems such as incomplete annotated data and low image resolution, making it difficult to train deep learning models with high recognition accuracy. Therefore, this article focuses on the research needs of deep learning based fine recognition tasks and constructs a high-resolution bird image dataset. The dataset contains a total of 200 classes of different birds, with 60 images for each class, totaling 12000 bird images. Each image in the dataset provides bird class information, three semantic parts bounding boxes, bird object segmentation masks, as well as the external characteristics, habitat environment, and geographic distribution of each bird class. This dataset was evaluated through fine-grained bird image classification, fine-grained bird image parts detection, and fine-grained bird image segmentation, verifying its effectiveness in fine-grained bird image recognition tasks. In the validation of the classification, the top-1 accuracies of SPDA-CNN, PS-CNN, MASK-CNN, Attention-CNN, Bilinear-CNN, RA-CNN are 90.7%, 81.2%, 91.4%, 83.6%, 89.3% and 91.6%. In the validation of bird parts detection, using three YOLOv5 models, the mAP of bird object are 83.52%, 85.35%, 85.96%, and the mAP of bird's head are 57.4%, 58.71%, 59.53%, and the mAP of bird's torso are 61.27%, 62.05%, 62.83%. In the validation of bird semantic segmentation, using ResNet-50 as a pre-trained encoder, the IoU of U-Net, PSPNet, and DeepLabV3+ is 84.98%, 89.69%, and 89.87%, respectively, verifying the effectiveness of this dataset in fine-grained bird image recognition tasks. The collected bird images and all annotated data involved are open source on the project homepage at https://www.ynbirds.cn/.