VisDrone Dataset Engineering & Benchmark Guide

Problem definition

How do I format, evaluate, and train aerial object detection models on the VisDrone dataset using reproducible annotations and verified class mapping?

Steps

1. Download Official VisDrone Subsets

   Fetch VisDrone2019-DET-train and VisDrone2019-DET-val archives from the official GitHub / repository sources.

2. Convert Bounding Box Annotations

   Convert VisDrone text format (bbox_left, bbox_top, bbox_width, bbox_height, score, object_category, truncation, occlusion) into normalized YOLO coordinates (class_id, x_center, y_center, width, height).

3. Map Class Indices

   Map the 10 object classes (pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor) explicitly in your dataset.yaml descriptor.

4. Execute Baseline Validation

   Train a baseline model using Ultralytics YOLO and record [email protected] and [email protected]:0.95 metrics against the evaluation split.

Guide

Judgment criteria

• Annotation coordinates correctly normalize to range [0, 1] matching image resolution.
• Validation dataset classes match official VisDrone category IDs exactly.

Common risks

• Ignored category 0 (ignored regions) being treated as a valid bounding box class.
• Misinterpreting occlusion and truncation flags causing false negative background masks.

Checklist