Close

Advanced School for Computing and Imaging (ASCI)

ASCI office
Delft University of Technology
Building 28, room E.3080
Van Mourik Broekmanweg 6
2628 XE – DELFT, The Netherlands

E: asci-office@tudelft.nl

Directions

The ASCI office is located at the Delft University of Technology campus.  It is easily accessible by bicycle, public transport and car. The numbers of buildings can help you find your way around the campus. Make sure you remember the name and building number of your destination.

Contact us at +31 15 278 8032 or send us an email at asci-office@tudelft.nl

Outdoor Scene Understanding under Adverse Weather Conditions’

Outdoor Scene Understanding under Adverse Weather Conditions’

Author : Qi Bi
Promotor(s) : Prof.dr. Th. Gevers / Dr. S. You
University : UVA
Year of publication : 2025
Link to repository : Link to thesis

Abstract

Understanding the outdoor scenarios is particularly challenging due to the various adverse weather conditions, which can degrade the scene representation greatly. This doctoral thesis focuses on understanding outdoor scenes under adverse weather conditions from both a physics-based vision perspective and a generalizable machine learning perspective. Five key research questions are studied.

(1) Weather can be modeled as a transitional state, allowing for the coexistence of multiple weather conditions. Weather uncertainty is introduced, considering both the probability levels and coexistence of multiple weather conditions. A physics-inspired prior-posterior weather uncertainty modeling scheme is proposed to advance this area.

(2) Domain-generalized urban-scene semantic segmentation aims to learn generalized semantic predictions across diverse urban-scene styles. A content-enhanced mask attention mechanism is proposed, which leverages the stronger context representation ability, while substantially maintaining the content information under cross-domain variation.

(3) The fog condition can pose severe occlusion to the scene object. This thesis proposes to learn foggy-scene segmentation under the domain generalization setting, which does not involve any foggy images in the training stage. A bi-directional wavelet guidance (BWG) mechanism is proposed.

(4) The generalized foggy-scene semantic segmentation is further studied from a perspective of physical based vision. A cross-domain foggy scene formulation model, along with a frequency decoupling scheme, is proposed.

(5) Existing semantic segmentation methods primarily focus on well-lit daytime scenarios and are not adequately designed to handle substantial appearance changes. A novel interactive learning scheme is proposed to understand the shift in illumination throughout the day affects scene semantics.