Tianyi Guan

About

This project investigated weakly supervised object recognition under partial occlusion and proposed a Bag-of-Alphabets (BOA) representation for recognizing overlapped objects. Rather than representing an object only through whole-object visual words, BOA decomposes recognition into local part-level "alphabets" that can remain detectable when other regions of the object are obscured. The framework combines local feature extraction, scale-space keypoint detection, feature filtering, orientation-aware descriptors, normalization, and clustering to construct an occlusion-tolerant visual representation.

Research Focus

To determine whether part-level visual representations can preserve object identity under substantial overlap, where conventional Bag-of-Words representations may fail because a large fraction of the object is no longer visible.

Study Design

The proposed BOA framework was developed from the Bag-of-Words/Bag-of-Features paradigm using SIFT-based local feature extraction, Gaussian and Difference-of-Gaussian scale-space analysis, unstable keypoint removal, orientation assignment, feature normalization, and K-means clustering. BOA and conventional BOW representations were then compared across 18 object categories under both non-occluded and occluded conditions, including humans, faces, animals, vehicles, food, landscape elements, and personal objects.

Key Findings

Under non-occluded conditions, BOA and BOW produced broadly comparable recognition performance. Under object overlap, however, the conventional BOW representation deteriorated substantially, whereas BOA retained strong recognition across all evaluated categories. Across the 18 reported occluded-object comparisons, the mean recognition score increased from approximately 0.209 with BOW to 0.889 with BOA, supporting the use of local part-level representations for occlusion-robust recognition.

Key Findings

Weakly supervised learning; SIFT feature extraction; Gaussian scale space; Difference-of-Gaussian keypoint detection; feature-point filtering; orientation descriptors; feature normalization; K-means clustering; Bag-of-Words/Bag-of-Features representation; Bag-of-Alphabets modelling.

Keywords

Computer vision; weakly supervised learning; object recognition; object occlusion; overlapped objects; Bag-of-Alphabets; Bag-of-Words; Bag-of-Features; SIFT; K-means; visual vocabulary; local feature representation;