DRASHTI-HaOBB: Drone nadiR-view Annotated imageS of veHicles dataseT for India - Heading-angle Oriented Bounding Box: Indian Vehicle Oriented-Object-Detection Dataset with 1.3 Million Samples

Dataset structure:

After downloading all six ZIP files, the dataset must be reorganised into the directory format described below. Each image directory contains JPEG images, and the corresponding label directory contains annotations in a text files.

DRASHTI-HaOBB/

• images/
  • train/ *.jpg
  • val/ *.jpg
  • test/ *.jpg
• labels/
  • train_original/ *.txt
  • val_original/ *.txt
  • test_original/ *.txt

Description of the label (txt) file:

• The first line in the txt file specifies the flight height of the drone for the particular image
• Each of the remaining lines in the txt file has the following info in that order:
  • x1, y1, x2, y2, x3, y3, x4, y4, class-label, difficulty-level, heading-angle
• The four points of the Oriented Bounding Box (polygon) of the vehicle are given by x1, y1, x2, y2, x3, y3, x4, y4
• A class-label could be one of the following 14 classes: ‘Auto3WCargo’, ‘AutoRicksaw’, ‘Bus’, ‘Container’, ‘Mixer’, ‘MotorCycle’, ‘PickUp’, ‘SUV’, ‘Sedan’, ‘Tanker’, ‘Tipper’, ‘Trailer’, ‘Truck’, ‘Van’
• A difficulty level could have either 0 or 1
  • 1 would indicate a cropped vehicle on the boundary of the image (indicating that it is hard to figure out the class of the vehicle, because it is only partially visible in the image)
• Heading-angle (in the range 0 to 359 degrees) provides the clockwise angle (w.r.t to the x-axis of the image coordinate system) of the vehicle's front (indicating the direction of motion of the vehicle) 

Description of the meta-level CSV file (DRASHTI-HaOBB_framewise_info.csv):

• Each row contains the following info:
  • Image name, Split (Image is part of which set: training/test/validation), Image is real or augmented (Augmented: Yes/No), and the number of vehicle samples in the Image under each of the 14 classes

List of known issues/observations:

• The dataset obviously consists of cropped vehicle samples (on the boundary of the image) because these vehicles are entering/leaving the scene. 
  • Such severely cropped vehicles on the boundary of the image, having one of the HBB (for the OBB) dimensions smaller than ~20 pixels, are not annotated, because they do not contain any usable visible information
  • For a few such cropped vehicles, especially on the boundary of the image, the OBB may not be a rectangle (but would still be a 4-point polygon)
• In the Auto3WCargo class, there are very few samples (~200) of non-motorised tricycles; This particular object was annotated/included because its cargo compartment is large and looks like an Auto3WCargo
• Sometimes, AutoRicksaw are carrying cargo-like objects on the top of the vehicle, and may appear like Auto3WCargo. Such vehicles are correctly classified as AutoRicksaw in the annotations
• In the Trailer class, there are a few samples (~2500) of agriculture tractor-trailer; There are also ~50 samples of only agriculture tractor
• Very few vehicles may be occluded by trees or by traffic light poles
• Few images are under smog and dusty locations (and the vehicles appear blurry), especially near Ready-Mix Cement sites, and sand/mud-filling stations
• The heading angle for parked vehicles is approximate
• The heading angle for the vehicle just entering / about to leave the image (scene) is approximate
• For the MotorCycle class, a very small percentage of bounding boxes would be misaligned; Also, sometimes, pedestrians with nearby shadows may look like a MotorCycle, and would have been inadvertently annotated as a MotorCycle
• Pedestrians and non-motorised objects (example: bicycles or tricycles) are not annotated
• During the final visual validation of the dataset using the tool Visual Similarity Duplicate Image Finder, we found total errors to be extremely small percentage (0.038 %) [~500 samples out of total 1.3 million samples],  which include (human errors): misaligned (or lagging) OBBs for Truck, SUV, AutoRicksaw and MotorCylce; un-annotated MotorCycle; empty OBBs for MotorCycle;