Tsinghua Dogs Dataset

Ding-Nan Zou^1,2, Song-Hai Zhang¹, Tai-Jiang Mu¹, Ming Zhang³

¹Tsinghua University, ²NaJiu Company, ³Harvard Medical School

Fig. 1 Variation in Tsinghua Dogs dataset. (a) Great Danes exhibit large variations in appearance, while (b) Norwich terriers and (c) Australian terriers are quite similar to each other.

Introduction

Tsinghua Dogs is a fine-grained classification dataset for dogs, over 65% of whose images are collected from people's real life. Each dog breed in the dataset contains at least 200 images and a maximum of 7,449 images, basically in proportion to their frequency of occurrence in China, so it significantly increases the diversity for each breed over existing dataset (see Fig. 1). Furthermore, Tsinghua Dogs annotated bounding boxes of the dog’s whole body and head in each image (see Fig. 2), which can be used for supervising the training of learning algorithms as well as testing them.

Fig. 2 Annoations: bounding boxes for whole dogs (blue) and their heads (red).

Following is the brief information about the dataset:

Number of categories: 130;
Number of images: 70428;
Annotations: class labels, bounding box of the dog head and bounding boxes of the whole dog body.

Download

The dataset provides two versions of images to download: high resolution and low resolution. Details about Tsinghua Dogs can be found in this paper [link or PDF].

Item	Download Link	Size
Low resolution images	low-resolution	2.5GB
Annotations for low resolution images	low-annotations	36MB
High resolution images	high-resolution.001	38.8GB
	high-resolution.002
	high-resolution.003
	high-resolution.004
	high-resolution.005
	high-resolution.006
	high-resolution.007
	high-resolution.008
	high-resolution.009
Annotations for high resolution images	high-annotations	36MB
Train and validation splits	TrainValSplit	0.2MB

Benchmarking

We have also benchmarked several classification methods on our dataset, including both general neural networks and fine-grained models which exhibit good performance on other fine-grained datasets.

Rank	Model	Backbone	Batchsize	Epochs	Accuracy(%)	Year
1	WS-DAN [code]	Inception v3	12	80	86.4	2019
2	TBMSL-Net [code]	Resnet50	6	200	83.7	2020
3	PMG [code]	Resnet50	16	200	83.5	2020
4	Inception v3 [code]	N/A	64	200	77.7	2016

Citing Tsinghua Dogs

Please cite our Tsinghua Dogs in your publications if it helps your research:

        @article{Zou2020ThuDogs,
          title={A new dataset of dog breed images and a benchmark for fine-grained classification},
          author={Zou, Ding-Nan and Zhang, Song-Hai and Mu, Tai-Jiang and Zhang, Min},
          journal={Computational Visual Media},
          year={2020},
          url={https://doi.org/10.1007/s41095-020-0184-6}
        }