Egocentric recognition of handled objects: Benchmark and analysis

Abstract

Recognizing objects being manipulated in hands can provide essential information about a person's activities and have far-reaching impacts on the application of vision in everyday life. The egocentric viewpoint from a wearable camera has unique advantages in recognizing handled objects, such as having a close view and seeing objects in their natural positions. We collect a comprehensive dataset and analyze the feasibilities and challenges of the egocentric recognition of handled objects. We use a lapel-worn camera and record uncompressed video streams as human subjects manipulate objects in daily activities. We use 42 day-to-day objects that vary in size, shape, color and textureness. 10 video sequences are shot for each object under different illuminations and backgrounds. We use this dataset and a SIFT-based recognition system to analyze and quantitatively characterize the main challenges in egocentric object recognition, such as motion blur and hand occlusion, along with its unique constraints, such as hand color, location prior and temporal consistency. SIFT-based recognition has an average recognition rate of 12%, and reaches 20% through enforcing temporal consistency. We use simulations to estimate the upper bound for SIFT-based recognition at 64%, the loss of accuracy due to background clutter at 20%, and that of hand occlusion at 13%. Our quantitative evaluations show that the egocentric recognition of handled objects is a challenging but feasible problem with many unique characteristics and many opportunities for future research.