Deng Yubin has won the ASM Technology Award 2015 with his undergraduate Final Year Project titled "Attribute based human identification" supervised by Prof. LOY Chen Change Cavan. His project is considered for the gold award as a practical innvoation tool and advanced methodology that are specific for pedestrian attribute recognition at far distance, and he will be awarded HKD50,000 as an encouragement to achieve a greater success. Yubin is currently a Year-1 PhD student from the Department of Information Engineering. About the winning project: Attribute based human identification Learning to recognize pedestrian attributes, such as gender, age, clothing style, has received growing attention in computer vision research, due to its high application potential in areas such as video-based business intelligence and visual surveillance. In real-world video surveillance scenarios, clear close-shots of face and body regions are seldom available. Thus, attribute recognition has to be performed at far distance using pedestrian full-body appearance in the absence of critical face/close-shot body visual information. In view of the lack of close-shot images, a novel approach is proposed in this study for far-view attribute recognition with emphasis on mitigating the visual ambiguity of appearance features. Specifically, we consider inference with the help from neighboring pedestrian images whose appearances look alike and we hypothesize that neighboring samples share natural invariance in their feature space. To this end, we view multiple pedestrian images as forming a Markov Random Field (MRF) graph with node associations weighted by pairwise image similarity. By carrying out inference on the graph, we jointly estimate the attribute probability of all images in the graph. Satisfactory attribute recognition performance on most attributes is reported on our newly released PETA dataset -- our new model surpasses state-of-the-art performance in far-view attribute recognition. |
|
