Seeking novel proposals able to automatically generate realistic image and video data to be used in Computer Vision and Machine Learning AI algorithm development

Background

Challenge

There is a high demand internally to collect large amounts of video and image data, which is currently manually carried out, to be used for AI algorithm development. There is a lack of suitable and sufficient datasets from an automotive environment, both inside and outside the car.

Current Existing Method

Collection of video recording data in vehicle is carried out manually.

The manual collection data has certain restriction and gaps to be addressed, e.g.

• Inability to cover dangerous driving scenario on the road.
• Inability to cover all possible cases and combinations of driving behaviour and activities.
• Lack of flexibility (different camera position, field of view, colour/Infrared (IR) camera types, etc)

Requirements

• The proposed software solution should generate realistic image and video data for in-vehicle view of drivers, covering different relevant human behaviour inside the car cabin environment. For example, a driver making a phone call, drinking, eating, smoking, having an object in hand, head and body movements, hand gestures, etc.
• An example of a solution we are seeking is to replace a person’s face in video with someone else’s via deepfake technology. This is to be able to automatically generate multiple images and videos of individuals with different facial features, exhibiting a range of behaviours, and carrying out various actions.
• The types of camera views we are looking at that are applicable to deepfake technology are those from RGB and infrared (greyscale) cameras, with resolutions of 1280x800 and 1600x1300. The common viewing angle is the front and side (45 degree angle) view of the driver.
• The solution should be highly flexible and configurable, to be able to generate the variety of in-vehicle view scenarios specified above.
• The operating and verification environment involves hardware or software in the loop testing.

There are already known solutions such as synthetic data generation to address certain use cases. Nevertheless, there is still further potential make these generated data more realistic with different techniques.

Ground truth image and video data captured by cameras may be provided to the solution provider for evaluation if required. Potential solution provider should be able to demonstrate a good concept with video created based on in-vehicle video recorded on their own or those downloaded from the Internet. Specific sample videos and ground truth data may be provided by Continental (with NDA) for further evaluation at a later stage.

Geographical Restrictions

Usage of Deepfakes could be restricted at different region due to legal law.

Minimum Required Technology Readiness Level (TRL)

Level 7

Development Timeframe

6 months

Trade and Connectivity Challenge 2020

The present Innovation Call is part of the Trade and Connectivity Challenge 2020 co-organised by Enterprise Singapore along with IPI Singapore.

Singapore’s position as a strategic trade hub is backed by its strong regional networks and connectivity. In recent years, Singapore has also transformed into a hub for innovation activity.

The Trade and Connectivity Challenge (TCC) leverages Singapore’s vibrant ecosystem of global trade, connectivity and innovation as a call for solutions to drive partnerships in the aviation, maritime, land mobility, logistics, and trade sectors. Now, in its second year, TCC 2020 continues to provide opportunities for startups and SMEs to co-innovate, deploy, and adopt innovative solutions.

Enterprise Singapore, together with IPI, invites you to showcase your innovative solutions in growth areas, such as big data analysis, Internet of Things and sustainability post-COVID19.