【Vision in Action】Beyond Strong Baseline - Training Data

🏁 Beyond Strong Baseline: Multi-UAV Tracking Competition ₊˚⊹

🎯 COMPETITION LINK: Codabench

⬇️ DOWNLOAD: Training Data | Test Data (BB2P_02_corrected) | Sample Submission

📹 VIDEOS: Overview | Training Videos | Test Videos

This project provides the training data for the upcoming multi-UAV tracking evaluation event.

The dataset has been carefully curated from 200 publicly released training videos.

To ensure fairness and minimize the risk of information leakage into the evaluation phase, we:

Renamed, resplit, and resampled the original videos. 

Selected 102 videos to serve as the training data.

Takeoff - UAV launch phase: 1 video.

L - Larger UAV target: 8 videos.

C - Cloud background: 20 videos.

CF - Cloud (Fewer UAVs): 9 videos.

T - Tree background: 34 videos.

TF - Tree (Fewer UAVs): 7 videos.

B - Scene with buildings: 6 videos.

BB1 - Building Background 1: 2 videos.

BB2 - Building Background 2: 9 videos.

BB2P - Building Background 2 (UAV partially out of view): 4 videos.

Landing - UAV landing phase: 2 videos.

Ensured that both the training and test sets are balanced across all scenario types to support robust evaluation.

The total number of frames across training sequences, whether from images or videos, is 77,293.

⚠️ Important Guidelines: Participants must train their models using only this dataset. Using external data may lead to overfitting or unreliable evaluation results.

📁 Folder Structure: The dataset is organized into three folders as below.

.├── TrainImages│       └── B_00│                 ├── 00000.jpg│                 ├── 00001.jpg│                 ├── 00002.jpg│                 └── ...├── TrainLabels│       ├── B_00.txt│       └── ...└── TrainVideos          ├── B_00.mp4          └── ...

📂 Data Format: The data contains:

.mp4: Raw videos for each sequence.

.jpg: Extracted frames from each video.

.txt: Frame-by-frame object annotations in MOT format.

Each .txt file contains annotations in the following format:

frame_id, object_id, x1, y1, w, h, confidence, class, visibility

frame_id: Index of the frame in the video (starting from 1).

object_id: Unique ID assigned to each object across frames.

x1, y1: Top-left coordinates of the bounding box.

w, h: Width and height of the bounding box.

confidence: Confidence score.

class: Integer label for the object class.

visibility: Visibility ratio.

Example:

1,1,330.31,10.11,9.050000000000011,5.59,1,1,1.0

1,2,360.93,13.96,9.370000000000005,5.34,1,1,1.0

🙌 Feel free to reach out in the GitHub Issues section if you have any questions. Let us improve the tracking performance together. Have fun!