Feature Clustering For Recurring Pattern Detection
Open Access
- Author:
- Zhang, Hong
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- July 09, 2021
- Committee Members:
- Chitaranjan Das, Program Head/Chair
Yanxi Liu, Thesis Advisor/Co-Advisor
Robert Collins, Committee Member - Keywords:
- recurring pattern detection
hierarchical clustering method
vanishing point detection
deep feature clustering - Abstract:
- Recurring pattern (RP) detection is one of the most challenging research topics in computer vision. We follow the definition of a recurring pattern in "GRASP Recurring Patterns from a Single View" (GRASP-RP, our baseline), where a recurring pattern refers to an ensemble of two or more instances with some shared visual features. Here, an instance may or may not correspond to a physical object. Tasks addressed by this thesis include: (1) propose an alternative hierarchical clustering with voting (HCV) algorithm to group similar visual features; (2) apply cross-ratio constraint from projective geometry on projective-view images for translation symmetry, and vanishing point discovery; and (3) use a pre-trained convolutional neural network (CNN) to extract deep features from patches and cluster deep features by HCV. We validate our methods on a newly labeled recurring pattern dataset with 1006 images, including both frontal and projective-view images. Both qualitative and quantitative results are presented. A two-level evaluation is presented: RP-level and RP-instance-level. Our results show that (1) At the RP-level: HCV with the DBSCAN feature filtering before GRASP-RP algorithm (DBHCV_GRASP) outperforms the baseline by 9% average precision (pvalue = 0.0021). The baseline outperforms DBHCV_GRASP by 3% average recall (pvalue = 0.0007). At the RP-instance level: DBHCV_GRASP outperforms the baseline by 3% mean average recall (pvalue = 0), while the baseline outperforms DBHCV_GRASP by 2% mean average precision (pvalue = 0). (2) The proposed method for vanishing point (VP) detection outperforms state of art Zhou’s method by 13% in terms of the mean success rate of vanishing point detection (pvalue = 0). (3) Finally, the deep feature clustering based on AlexNet (DFHCV_AlexNet) [5] outperforms the baseline by 69% RP-level recall (pvalue = 0). The baseline outperforms the DFHCV_AlexNet by 4% RP-level precision (pvalue = 0.0777). At the RP-instance level: we take the best-fit detected RPs of all the methods for comparison, the baseline achieves the highest mean average precision over DFHCV_AlexNet by 17% (pvalue = 0), while DFHCV_AlexNet outperforms the baseline by 11% mean average recall(pvalue = 0). In summary, for task (1), DBHCV_GRASP has improved the baseline on RP-level precision but not the RP-level recall. Also, it improves the baseline on RP-instance- level average recall but not the precision. For task (2), the proposed method outperforms Zhou’s VP detection method significantly. For task (3), our proposed method based on deep features has improved recall rates at both RP and RP-instance levels significantly, but not for precision rates.