Fish Stereo Matching using Modified k-Dimensional Tree Nearest Neighbor Search

Stereo matching is one of the primary determining factors of 3D reconstruction accuracy. This study proposes a modified k-d tree algorithm to match the detected centroids of the fish in 527 pairs of stereo images. The primary purpose of the modified k-d tree is to eliminate the matching errors caused by the close proximities of the fish as measured by the convex hull. With the fish proximities at not greater than 500 convex hulls, the probability of successful matching is only 5%. Considering all the input image pairs with convex hulls of 0 to 6500, the centroid matching using the conventional k-d tree obtained the maximum precision of S7%. With the implementation of the modified k-d tree, the matching precision has increased to 100%, which means the errors in centroid matching were totally eliminated with all the varying levels of fish proximities.