TrackingNet Large-Scale Outdoor Object Tracking Dataset

TrackingNet is a large-scale object tracking dataset consisting of outdoor videos. The dataset contains 30,643 video clips, divided into 30,132 training videos and 511 test videos, with an average of 470,9 frames per video.

The dataset provides more than 14 million dense bounding box annotations, covering various situations in outdoor scenes, including various frame rates, resolutions, context scenes and target categories.

Different from previous target tracking datasets, TrackingNet is divided into two parts: training set and test set. The authors carefully selected 30,132 training videos from Youtube-Bounding Boxes and constructed 511 videos with similar distribution to the training set to form a test set.