1. 將Query的圖分別利用HKM及LSH做 search的動作，分別得到一群與Query相似的圖 set(也許取top N)，而且需要記錄Result set中每個Feature 對應到 Query image中的哪個 Feature。
2.接著將兩個方法所找出來的Features對應關係合併(同張圖兩種方法可能會有不同的對應，取聯集)。
3.然後將聯集過後的圖 set，分別與Query image 去做 spatial verification (RANSAC) 依能對應到的最多inliers的多少排序，並由最高的Region(排序是以Region為單位)依序當成Query再次丟入兩邊方法。
4.將步驟3得到的結果重覆步驟2直到不再能找到新的Region。
2008年12月7日 星期日
more detail of our final project
Query expansion is a way to get the better search results from take the results of original query as new queries. And from the paper "Total Recall: Automatic Query Expansion with a Generative Feature Model for Object Retrieval", it uses five different algorithm for query expansion, and evaluating the performance by the Oxford dataset.
For image search, feature extraction is a very important part. The latest few years a keypoint based feature presentation called "visual word" have been used and having better performance in object detection then other feature presentation.
To decide the "visual word", clustering method like "k-means" is the simplest way. But the training phase often needs long time, and we have observed that a hierarchical k-means(HKM) and locality sensitive hash(LSH) maybe could improve the efficient of training phase, and the two methods have some fundamental differences. So, we think that maybe we can use them in the query expansion to experiment that could one method can get better search result by using the search results by the combination of two result as new queries.
What we plan to do is the following:
1. training the visual word by HKM and LSH.
2. query image as a input, search for the visual word both method generated.
3. using the combination of both result lists for query expansion input for both method.
4. implementing the 5 query expansion method discussed in the paper and evaluating the performance.
For image search, feature extraction is a very important part. The latest few years a keypoint based feature presentation called "visual word" have been used and having better performance in object detection then other feature presentation.
To decide the "visual word", clustering method like "k-means" is the simplest way. But the training phase often needs long time, and we have observed that a hierarchical k-means(HKM) and locality sensitive hash(LSH) maybe could improve the efficient of training phase, and the two methods have some fundamental differences. So, we think that maybe we can use them in the query expansion to experiment that could one method can get better search result by using the search results by the combination of two result as new queries.
What we plan to do is the following:
1. training the visual word by HKM and LSH.
2. query image as a input, search for the visual word both method generated.
3. using the combination of both result lists for query expansion input for both method.
4. implementing the 5 query expansion method discussed in the paper and evaluating the performance.
訂閱：
文章 (Atom)