Joint Segmentation and Recognition of Categorized Objects from Noisy Web Image Collection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing object segmentation methods often assume noise-free image collections.
• Web-scraped image collections frequently contain irrelevant or noisy images.
• This limitation hinders the performance of current categorized object segmentation techniques.

Purpose of the Study:

• To develop a method for automatic segmentation and recognition of categorized objects from noisy web image collections.
• To overcome the limitations of existing methods that fail with non-target object images.
• To enable accurate object segmentation even when the input collection is not curated.

Main Methods:

• A co-training framework combining object segmentation and category recognition algorithms.
• Object segmentation algorithm trained on high-confidence target object images.
• Category recognition model guided by intermediate segmentation results.
• Automatic identification of true positive images and extraction of target objects.

Main Results:

• The proposed co-training method effectively segments and recognizes categorized objects from noisy datasets.
• Demonstrated superior performance compared to state-of-the-art methods across four diverse datasets.
• Validated efficacy on datasets including the Weizmann horse, MSRC, iCoseg, and a new 30-category web image dataset.

Conclusions:

• The co-training approach successfully addresses the challenge of segmenting categorized objects in noisy web image collections.
• The method exhibits robustness and adaptability to real-world, uncurated image data.
• This work advances the field of object segmentation by enabling reliable processing of imperfect image datasets.