25,000 high-resolution images (split into 18,000 for training, 2,000 for validation, 5,000 for testing; at an average resolution of ~9 megapixels) with pixel-wise annotations from >2 million manually drawn polygons
100 object categories, 60 of those instance-specific (i.e. to enumerate objects)
Global geographic reach, covering North and South America, Europe, Africa, Asia, and Oceania
High variability in weather conditions (sun, rain, snow, fog, haze) and capturing times (dawn, daylight, dusk, and even night)
Broad range of camera sensors, varying focal length, image aspect ratios, and different types of camera noise
Different capturing viewpoints (from road, sidewalks, and off-road)
The public set comprises 20,000 images, out of which 18,000 shall be used for training and the remaining 2,000 for validation. The official test set contains 5,000 RGB images, which will be released separately. We provide pixel-wise labels based on polygon annotations for 66 object classes, where 37 are annotated in an instance-specific manner (i.e. individual instances are labeled separately). The folder structures contain raw RGB images ({training,validation}/images), class-specific labels for semantic segmentation (8-bit) ({training,validation}/labels) and instance-specific annotations (16-bit) ({training,validation}/instances). Please run ‘python demo.py’ from the extracted folder to get an idea about how to access label information and for mappings between label IDs and category names.
Bird ( 0): animal--bird has instances: True Ground Animal ( 1): animal--ground-animal has instances: True Curb ( 2): construction--barrier--curb has instances: False Fence ( 3): construction--barrier--fence has instances: False Guard Rail ( 4): construction--barrier--guard-rail has instances: False Barrier ( 5): construction--barrier--other-barrier has instances: False Wall ( 6): construction--barrier--wall has instances: False Bike Lane ( 7): construction--flat--bike-lane has instances: False Crosswalk - Plain ( 8): construction--flat--crosswalk-plain has instances: True Curb Cut ( 9): construction--flat--curb-cut has instances: False Parking (10): construction--flat--parking has instances: False Pedestrian Area (11): construction--flat--pedestrian-area has instances: False Rail Track (12): construction--flat--rail-track has instances: False Road (13): construction--flat--road has instances: False Service Lane (14): construction--flat--service-lane has instances: False Sidewalk (15): construction--flat--sidewalk has instances: False Bridge (16): construction--structure--bridge has instances: False Building (17): construction--structure--building has instances: False Tunnel (18): construction--structure--tunnel has instances: False Person (19): human--person has instances: True Bicyclist (20): human--rider--bicyclist has instances: True Motorcyclist (21): human--rider--motorcyclist has instances: True Other Rider (22): human--rider--other-rider has instances: True Lane Marking - Crosswalk (23): marking--crosswalk-zebra has instances: True Lane Marking - General (24): marking--general has instances: False Mountain (25): nature--mountain has instances: False Sand (26): nature--sand has instances: False Sky (27): nature--sky has instances: False Snow (28): nature--snow has instances: False Terrain (29): nature--terrain has instances: False Vegetation (30): nature--vegetation has instances: False Water (31): nature--water has instances: False Banner (32): object--banner has instances: True Bench (33): object--bench has instances: True Bike Rack (34): object--bike-rack has instances: True Billboard (35): object--billboard has instances: True Catch Basin (36): object--catch-basin has instances: True CCTV Camera (37): object--cctv-camera has instances: True Fire Hydrant (38): object--fire-hydrant has instances: True Junction Box (39): object--junction-box has instances: True Mailbox (40): object--mailbox has instances: True Manhole (41): object--manhole has instances: True Phone Booth (42): object--phone-booth has instances: True Pothole (43): object--pothole has instances: False Street Light (44): object--street-light has instances: True Pole (45): object--support--pole has instances: True Traffic Sign Frame (46): object--support--traffic-sign-frame has instances: True Utility Pole (47): object--support--utility-pole has instances: True Traffic Light (48): object--traffic-light has instances: True Traffic Sign (Back) (49): object--traffic-sign--back has instances: True Traffic Sign (Front) (50): object--traffic-sign--front has instances: True Trash Can (51): object--trash-can has instances: True Bicycle (52): object--vehicle--bicycle has instances: True Boat (53): object--vehicle--boat has instances: True Bus (54): object--vehicle--bus has instances: True Car (55): object--vehicle--car has instances: True Caravan (56): object--vehicle--caravan has instances: True Motorcycle (57): object--vehicle--motorcycle has instances: True On Rails (58): object--vehicle--on-rails has instances: False Other Vehicle (59): object--vehicle--other-vehicle has instances: True Trailer (60): object--vehicle--trailer has instances: True Truck (61): object--vehicle--truck has instances: True Wheeled Slow (62): object--vehicle--wheeled-slow has instances: True Car Mount (63): void--car-mount has instances: False Ego Vehicle (64): void--ego-vehicle has instances: False Unlabeled (65): void--unlabeled has instances: False
