Topic Title:  Fundamental Deep Learning Network for City Objects

 

Technical Area:

Computer Vision, Deep Learning

 

Background

With the development of computer vision technology in recent years, the fundamental deep learning networks not only achieve state-of-the-art performance on the large-scale image classification task, but also expanded to various other visual recognition tasks like object detection and semantic segmentation.

However, the existing fundamental deep learning networks, which are designed for universal image recognition, are suboptimal in the complex, wild scenarios of smart city. The main challenges include but not limited to: the poor image quality of surveillance, the variable environment factors, e.g., weather, seasons and illumination, the fine-grained differences between identities.

As a result, a fundamental deep learning network specifically designed for city objects (people, vehicles, objects and events) becomes a major requirement. We are looking for the researchers who are experts either in this field or the related fields and also keen to aim for the challenges and opportunities on this topic.

 

Target

  1. Design a fundamental deep learning network for city objects. Combine the superiority of existing network blocks and related prior knowledge about the properties of city objects.
  2. The designed network can be utilized as the backbone model for a series of computer vision tasks, e.g., recognition, detection, tracking, segmentation, re-identification, action recognition, abnormal detection.
  3. The designed network should outperform the existing CNN model (e.g. ResNet, DenseNet) at the same grade of parameters and FLOPs.
  4. The interpretability of the designed network should be proven.

 

Related Research Topics

  1. Convolution neural network for large-scale image classification, e.g. AlexNet, GoogLeNet, VGG, ResNet, DenseNet.
  2. Some sub-problems related to this topic: object detection, visual tracking, person re-id, action recognition, abnormal detection.