Nature Dataset

Click on these links to download this data set:

Images: nature_frames
Label: labels

sample categories

Representative images for each category.


This data set captures the natural distribution of object appearance, disappearance, and interactions. While most objects have very few instances (< 50), there are a few objects that have a large number of instances (~ 800). The natural occurrence of nuisances is also present in this data set. Birds are cut off by the image frame, occlude one another, and get occluded by background objects. The background exhibits a high degree of variance due to lighting changes, even at the time-scale of a single hour. The background objects are not completely static either: feeders and leaves swing in the wind. The motion of the feeders are also affected by birds landing on the feeder posts.


Images are captured at a rate of one per second from a camera pointed at a feeder station in a natural reserve environment. The annotated portion consists of 3600 color images of 480x704 pixels. Each image is annotated with a bounding box enclosing each bird instance, including flags that indicate whether or not it was interlaced or occluded to the point where categorization could not be performed even by human experts. Each bounding box is also labeled with an object id and a category id. In this case, object refers to a unique bird and category refers to a bird species. We detected 7932 instances, organized into 358 objects and 17 different categories, where one category contains unidentifiable objects. From that, there are 5863 “good” instances that are not occluded or interlaced and 199 objects that have 3 or more “good” instances. The table below details how the objects and instances per object are distributed for each category.


1) T. Ko, S. Soatto, and D. Estrin. Background Subtraction on Distributions. In European Conference on Computer Vision, pages 222-230. [PDF]
2) T. Ko, S. Soatto, and D. Estrin. Categorization in Natural Time-Varying Image Sequences. To appear in Computer Vision Pattern Recognition: Visual Interpretation and Understanding Workshop, June, 2009. [PDF]

When publishing results using this data set, please cite (2). If you would like to have your paper appear on this list, please email me.