Visual Cue Cluster Construction via Information Bottleneck Principle and Kernel Density Estimation

Winston H. Hsu and Shih-Fu Chang

CIVR 2005 Reviews' comments :

	Reviewer 1	Reviewer 2	Reviewer 3
Relevance to CIVR	Very Significant	Very Significant	Very Significant
Significance	Significant	Significant	Significant
Clarity	Clear	Very Clear	Very Clear
Originality	Original	Original	Original
Correctness & Soundness	Mostly Correct & Sound	Correct & Sound	Correct & Sound
Comparison with Prior Work	Adequate	Very Adequate	Very Adequate
Overall Recommendation	Strong Accept	Strong Accept	Strong Accept
Reviewer's familiarity with topic and confidence with review	Very familiar & very confident	Very familiar & very confident	Very familiar & very confident
Comments to authors		This paper presents a novel visual cue cluster construction (VC^3) method to automatically discover mid-level features. This method is based on the information bottleneck principle. The authors use a large data set of videos (the standard Trec2004) and the results are convincing. The experimental results indicate the superiority of the proposed method, especially in the tested CNN videos. The results also clearly show that VC^3 outperform the K-means in the CNN videos.	This paper proposed essentially a new feature transformation framework, where raw features (e.g., image features) are converted to cue cluster soft membership values (probabilities). The cue clusters are learned based on the sIB algorithm proposed in [4-5] and used in [6]. sIB is a greedy algorithm for clustering with prior information (class labels). A SVM is used on these features to perform classification. The presentation is very clear. And the method and resutls are reasonable and convincing. Comments: - With a large number of training samples, Kernel Density Estimate (e.g., Eq. 7) can become forbiddenly expensive; - Text in Figures are messed up! - Typo: Section 2.4, till--> still