Indian Road Driver’s View (IRDV) Dataset

Adverse driving conditions like darkness, rain, and fog present significant challenges to professional drivers as well as to computer vision algorithms in autonomous vehicles. One potential solution is to use an on-board system for real-time image translation, transforming weather-affected images into clear ones.
While methods like autoencoder and pix2pixGAN are effective for similar translation problems, the lack of paired images for training makes it infeasible to train these models. This paper solves this issue by generating a synthetic dataset using a combination of physics-inspired translation and style transfer. The novelty lies in the physics models that transform images based on physical laws, while style transfer makes the output images look realistic.
An experiment was designed to create an image dataset of Indian road images under proper daylight conditions. This dataset is named the Indian Road Driver’s View (IRDV) Dataset.

Several locations with varying road types, such as urban, rural, highways, and environmental conditions, were chosen to capture a diverse range of road scenes. The cities selected to capture road images are Guwahati, Hyderabad, Kolkata, Chennai, Jaipur, Mumbai, Bengaluru, Gandhinagar, and New Delhi. experiment was conducted during clear daylight conditions to ensure proper lighting across all captured images. Smartphone cameras were mounted on the car bonnet. Auto settings of shutter speed and ISO were used to ensure well-exposed images without overexposure or underexposure issues. Careful attention was paid to the position and orientation of the camera relative to the sun’s direction to avoid glare and lens flares and enhance the clarity of the captured images. Videos were captured systematically to ensure coverage of different road types, traffic conditions, and lighting angles, and frames were extracted at intermittent intervals to ensure variations.