The full dataset has been released. There are three subsets in the dataset, i.e., the train subset, the test subset for track 1 and the test subset for track 2. The train subset consists of 200 thermal infrared video sequences and publishes detailed annotation files (whether the target exists, target location information and various challenges). The subset for track 1 also contains 200 video sequences, only providing the position information of target in the first frame; The subset for track 2 contains 200 video sequences. This track does not provide any labeled information. It requires participants to obtain the flag of existence and corresponding target location information of the target through detection and tracking. Above three subsets do not have any overlap between each other. We propose participants could train a suitable detector or tracker model depending on multiple categories of label information in train subset.
CVPR 2023 Anti-UAV Challenge Dataset
1. Folder Tree Diagram
Compared to the previous challenge, we further enlarge the dataset this year by adding more challenging video sequences with dynamic backgrounds, complex movements, and tiny-scale targets, such that the resulting new dataset covers a greater variety of scenarios with multi-scale UAVs.
Examples are shown as follows.
- cloud background
- building background
- mountain background
- sea background
- fast movement
- scale variation
- no-target scene
- large target
- medium target
- small target
- tiny target
3. Division Rules
train: 200 videos for training your algorithm.
test-track1: 200 videos for the final submission for track 1.
test-track2: 200 videos for the final submission for track 2.
4. Label Information
|OV/VE||Out-of-View:the target moves out of the current view.|
|OC||Occlusion: the target is partially or heavily occluded.|
|FM||Fast Motion: the target moves quickly.|
|SV||Scale Variation: the scale of the bounding boxes over the frames vary significantly.|
|TC/IC||Thermal/Infrared Crossover: the target has a similar temperature to other objects or background surroundings.|
|DBC||Dynamic Background Clusters: there are dynamic changes (e.g., waves, leaves, birds) in the background around the target.|
|LR||Low Resolution: the area of the bounding box is small.|
|TS||Target Scale: the target is with a tiny, small, medium or large scale.|
You can download train set at Modelscope now.