LTDv2: A Large-Scale Long-term Thermal Drift Dataset for Surveillance

LTDv2 is a single-scene thermal multi-object detection dataset created by vapaau. It contains over 1 million frames extracted from video clips spanning 8 months, with more than 6.8 million annotated bounding boxes for 4 object classes. The data encompasses several seasons, weather conditions, and day-night cycles.

Use Cases

• Training thermal object detection models based on the over 6.8 million annotated bounding boxes.
• Developing algorithms robust to thermal drift based on data spanning 8 months and multiple seasons.
• Evaluating model performance across varying conditions based on the described day-night cycles and weather variations.
• Benchmarking multi-class detection systems for surveillance based on the four object classes (Person, Bicycle, Motorcycle, Vehicle).

Strengths

• Large scale with over 1 million frames.
• Extensive annotations with more than 6.8 million bounding boxes.
• Long temporal coverage spanning 8 months and multiple seasons.
• Captures diverse conditions including day-night cycles and various weather.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is limited to a single scene, which may affect generalizability.

Provenance

Source

huggingface user vapaau

Collection Method

Extracted from video clips, likely from a fixed surveillance camera.

Time Range

Spans a period of 8 months, covering several seasons.

Freshness

Last updated 2025-10-15 21:25:55; freshness should be verified.