DhakaDrive: 2D Bounding Box Annotated Dashcam Videos of Unstructured Urban Traffic

Two weeks of dashcam video recordings from Dhaka, Bangladesh, captured under daytime and nighttime conditions. The dataset comprises 173 high-resolution videos totaling 6 hours and 18 minutes, with approximately 634,000 frames annotated with 2,180,000 bounding boxes across 12 distinct vehicle classes. It was created by Mohammed, Saif and last updated on March 29, 2026.

Use Cases

• Train object detection models based on 2D bounding box annotations for 12 vehicle classes.
• Benchmark detection performance on locally specific vehicle morphologies underrepresented in existing datasets.
• Research domain adaptation between structured (western) and unstructured (Global South) traffic environments.
• Evaluate model robustness under varying lighting conditions based on daytime and nighttime recordings.

Strengths

• Contains 173 high-resolution (2592 × 1944 pixels) video recordings totaling 6 hours and 18 minutes.
• Provides approximately 634,000 annotated frames with 2,180,000 bounding boxes.
• Covers 12 distinct transportation vehicle classes, including categories specific to Bangladesh.
• Includes standard train, validation, and test splits to facilitate reproducible experiments.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect geographic and temporal bias inherent to a two-week collection in Dhaka.

Provenance

Source

Harvard Dataverse

Collection Method

Frame-level 2D bounding box annotated dashcam video recordings collected from real-world urban traffic.

Time Range

Two weeks (specific dates unknown).

Freshness

Last updated 2026-03-29 14:26:53; freshness should be verified.

Geography

Dhaka, Bangladesh, covering North to South.