| dc.description.abstract | Autonomous Vehicles (AVs) are being partially deployed and tested across various global locations, including China, the USA, Germany, France, Japan, Korea, and the UK, but with limited demonstrations in Australia. The integration of machine learning (ML) into AV perception systems underscores the need for locally labeled datasets to develop and test algorithms in specific environments. To address this, we introduce SydneyScapes—a dataset tailored for computer vision tasks of image semantic, instance, and panoptic segmentation. This dataset, collected from Sydney and surrounding cities in New South Wales (NSW), Australia, consists of 756 images with high-quality pixel-level annotations. It is designed to assist AV industry and researchers by providing annotated data and tools for algorithm development, testing, and deployment in the Australian context. Additionally, we offer benchmarking results using state-of-the-art algorithms to establish reference points for future research and development. | en_AU |
| dc.description.method | Data collection, post-processing, and annotation of images for semantic, instance, and panoptic segmentation is an important step in training and testing a computer vision system to accurately recognise and classify objects in images from our local environment. This section explains the process from the sensor setup to the final labelled dataset.
We used a single camera mounted on an urban vehicle (Volkswagen Passat) for data collection. The camera, an SF3324 automotive GMSL model, has an ONSEMI CMOS Image Sensor AR0231 (2M Pixel) and a SEKONIX ultra-high-resolution lens. The lenses provide a 120-degree horizontal field of view (FOV) and a 73-degree vertical FOV. Images were captured at a resolution of 1928 x 1208 pixels (2.3M pixels).
The data collection was primarily conducted in Sydney, Australia, with some images also captured in nearby towns. The data was recorded naturally, with the vehicle being driven on the road under normal conditions. The focus was on capturing a variety of scenarios across different lighting conditions and diverse crowd densities. From multiple driving sessions, we selected a total of 755 images for labelling. These images represent different scenarios and provide coverage for training and evaluating segmentation models in Australian environments.
For visual consistency, we divided the dataset into three parts - SydneyScapes day, night, and people - to fine-tune ML methods on a diverse range of data, including both day and night conditions and different types of objects and scenes.
SydneyScapes day
The dataset includes 332 images of different cities around NSW, including Cudal, Orange, and Sydney. The data collection took place on rural roads, highways, and urban areas. This dataset comprises several conditions, such as sunny, cloudy, rainy, strong shadows, etc.
SydneyScapes night
Changes in illumination can challenge CV tasks by altering the appearance of objects, making recognition difficult. To address this, we recorded data at night to assess image segmentation algorithms under low-light conditions. The night dataset comprises 104 images collected in urban areas of Sydney.
SydneyScapes people
This dataset focuses on detecting and segmenting people in urban driving environments, an important aspect of AV safety systems. Accurate pedestrian detection is important for safe navigation, especially in high-density urban areas and zones with increased pedestrian presence. This dataset includes 320 images collected in different suburbs of Sydney.
Anonymisation
In compliance with local authority policies requiring data to be anonymised before publication, we developed a post-processing pipeline with two key algorithms: removing and replacing human faces and another blurring number plate information.
For face anonymisation, we used the DeepPrivacy algorithm, which anonymises faces by generating realistic, privacy-safe substitutes while preserving the original background of the image. To anonymise number plates, we employed the DashCamCleaner; this method uses the YoloV8 algorithm to detect the bounding boxes containing number plates. Then, it applied a Gaussian blur to the pixels within these bounding boxes.
Labelling
We drew inspiration from the Cityscapes dataset for labelling the images, adopting their labelling policy and original labels. We reorganised the class definitions into groups similar to those in Cityscapes. The dataset labels are divided into seven groups: Flat, Human, Vehicle, Construction, Object, Nature, and Void. Each group contains specific labels for various items. For instance, the "Vehicle" group includes labels for several types of transportation, such as "Car," "Bus," "On rails," "Motorcycle," "Bicycle," "Caravan," and "Trailer." Only the "Human" and "Vehicle" groups are classified as "things," for which we provide instance-level annotations. | en_AU |
