Multi-Task UAV Dataset for Floating Plastic Waste Detection in the Kashmir Region (AquaPlasticWaste 2026)
Description
AquaPlasticWaste 2026 — Multi-Task UAV Dataset for Floating Plastic Waste Detection in the Kashmir Region — is a 5,000-image UAV aerial imagery dataset for automated plastic waste monitoring on Dal Lake, an ecologically important Himalayan freshwater lake in Srinagar, Kashmir, India, facing severe pollution pressure. Images were captured with a consumer-grade quadrotor UAV flown at a consistent altitude across open water, shoreline, houseboat clusters, and floating garden zones, under varied lighting conditions. It is the first multi-task, polygon-annotated UAV dataset targeting plastic pollution on a Himalayan lake surface. All images carry polygon-based instance segmentation masks, rather than bounding boxes, to capture the irregular shapes of floating litter (e.g., flattened bags, torn wrappers). Masks were created in CVAT and verified through a primary annotator, a second cross-checking annotator, a third adjudicator, and a final quality-control pass. Two labeling schemes are provided from the same masks: binary (plastic vs. non-plastic) and seven-class (background, bottle, cap, foam, polythene, shoe, wrapper). Annotations are exported in both COCO JSON and YOLO polygon/segmentation formats for direct use in standard detection and segmentation frameworks. Of the 5,000 images, 4,692 (93.8%) contain plastic and 308 (6.2%) are plastic-free, with 30,166 polygon instances annotated in total. Wrapper is the most common class (35.4% of instances, present in 63.2% of images); shoe is rarest (1.1% of instances, 4.8% of images). Mean plastic pixel coverage is 1.40% per image (median 0.55%, max 33.31%), and median instance size is 1,448 pixels, well below typical benchmark object sizes (e.g., MS-COCO) — a genuine small-object challenge. Imbalance is reported at three levels for different loss-function needs: 31.9:1 by instance count (wrapper vs. shoe), 44.8:1 by pixel area among foreground classes (polythene vs. cap), and 70.4:1 between plastic and background pixels overall. A single annotation source supports four downstream tasks: image classification, litter-density regression, instance segmentation, and semantic segmentation, enabling multi-task and cross-task benchmarking without repeated labeling effort. Existing waste-detection datasets such as TrashNet, TACO, UAVVaste, and TrashCan are ground-level, underwater, or single-task; none target a freshwater lake surface from the air. AquaPlasticWaste 2026 fills this gap, offering a realistic, class-imbalanced, small-object benchmark relevant to environmental monitoring, lake-conservation planning, and edge-deployable UAV vision systems.
Files
Steps to reproduce
(1) Site Selection and Flight Planning. Map Lake into operational zones such as open water, shoreline, houseboat clusters, and floating gardens. Plan UAV flight paths covering all zones across multiple lighting conditions (morning, midday, overcast) to capture full environmental variability without bias toward any single zone. (2) Aerial Image Acquisition. Fly a consumer-grade quadrotor UAV with a high-resolution RGB camera at a fixed, consistent altitude to maintain uniform ground sampling distance. Conduct multiple flight sessions across different days and conditions to maximize environmental diversity. (3) Data Curation and Filtering. Review all captured frames and discard those that are blurred, overexposed (glare-affected), or near-duplicate. Retain only images of consistent, usable quality. This filtering reduced raw captures to the final retained set of 5,000 images. (4) Annotator Calibration. Assemble annotators familiar with expected litter types (bottles, bags, foam, footwear, wrappers). Run a calibration phase on a shared image subset so annotators agree on conventions for submerged items, overlapping instances, and glare-related boundary ambiguity before large-scale labeling. (5) Polygon-Based Annotation. Using CVAT, draw polygon masks (not bounding boxes) around every litter instance to preserve irregular shapes such as flattened bags or torn wrappers. Apply both labeling schemes from the same masks: binary (plastic/non-plastic) and seven-class (background, bottle, cap, foam, polythene, shoe, wrapper). (6) Cross-Verification. Each image is labeled independently by a primary annotator, then checked by a second annotator for boundary accuracy and category correctness. Disagreements are escalated to a third reviewer, who acts as final adjudicator. (7) Quality-Control Pass. Audit the full dataset to remove degenerate polygons, such as self-intersecting or near-zero-area masks, and correct remaining category errors. Measure the boundary anti-aliasing ratio to confirm the pipeline produces clean, unambiguous masks. (8) Format Export. Export finalized annotations into COCO JSON (detection/segmentation pipelines) and YOLO polygon/segmentation format (real-time frameworks like YOLOv8), both generated from the same annotation source to keep statistics consistent. (9) Dataset Splitting. Partition images into fixed train, validation, and test splits, avoiding near-duplicate leakage across splits, for standardized, comparable benchmarking. (10) Statistical Validation. Compute and cross-check class distribution, pixel coverage, instance density, object-size statistics, and inter-class co-occurrence across all export formats to confirm internal consistency before public release.
Institutions
- National Institute of Technology SrinagarJammu and Kashmir, Srinagar
Categories
Funders
- IIIT Hyderabad ihub-dataGrant ID: IIIT-H/IHub/Project Special Category_ST/2026-27/ST3-003