Vehicle Images for Binary Classification of Unsafe Passenger Hanging in Dhaka, Bangladesh

Description

This dataset contains 523 images for binary classification of unsafe passenger hanging in buses and legunas in Dhaka, Bangladesh. Positive means unsafe passenger hanging (251 images); Negative means safe/normal (272 images). These image-level class definitions were confirmed by the project representative during dataset preparation. The original hierarchy is preserved under Final_Dataset: Bus/Positive (164), Bus/Negative (178), Leguna/positive (87), and Leguna/negative (94). Normalize label capitalization when loading the data. Bus and Leguna identify vehicle groups; the prediction target has two classes. There are 404 HEIC, 41 JPEG, and 78 PNG files with variable dimensions. All 523 images passed full decoding checks. Original filenames, formats, folder names, and image bytes are retained. The file manifest provides image paths, vehicle groups, source labels, sizes, dimensions, and SHA-256 hashes. There are 522 distinct file hashes. Leguna/negative/IMG_3807.HEIC and IMG_3951.HEIC are byte-identical and both retained. No predefined training, validation, or test split is supplied. Keep duplicates and related capture sequences in the same partition and document preprocessing and split membership. Labels describe whole images; bounding-box annotations are not required for this classification task. Safe/normal is the assigned class, not a certification of overall transport safety.

Steps to reproduce

1. Download Final_Dataset.zip or the direct Final_Dataset folders, plus the documentation. These are two download options for the same 523 images; use one copy. Verify image paths and SHA-256 hashes against file_manifest.csv. 2. Assign the binary target from the source folder: Bus/Positive and Leguna/positive are positive (unsafe hanging); Bus/Negative and Leguna/negative are negative (safe/normal). A numeric encoding is positive=1, negative=0. Keep vehicle type as separate metadata. 3. Use a reader that supports HEIC, JPEG, PNG, and embedded orientation. Save conversions and resized images separately and document preprocessing. 4. Create training, validation, and test partitions before model development. Keep the exact duplicate pair and related capture sequences within the same partition. Account for class and vehicle distributions; no predefined split or acquisition-session grouping is supplied. 5. Fit and augment training data only. Use validation data for model selection and held-out test data for final evaluation. Report split membership, random seeds, preprocessing, exclusions, confusion matrix, and per-class precision, recall, and F1.

Institutions

• University of Dhaka

  Dhaka Division, Dhaka

Categories

Licence