Making specific features less discriminative to improve point-based 3D object recognition

Abstract

We present a framework that retains ambiguity in feature matching to increase the performance of 3D object recognition systems. Whereas previous systems removed ambiguous correspondences during matching, we show that ambiguity should be resolved during hypothesis testing and not at the matching phase. To preserve ambiguity during matching, we vector quantize and match model features in a hierarchical manner. This matching technique allows our system to be more robust to the distribution of model descriptors in feature space. We also show that we can address recognition under arbitrary viewpoint by using our framework to facilitate matching of additional features extracted from affine transformed model images. The evaluation of our algorithms in 3D object recognition is demonstrated on a difficult dataset of 620 images.

3D Object Recognition and Pose Estimation for Multiple Objects Using Multi-Prioritized RANSAC and Model Updating

“…In Table 1, each weight is evaluated in terms of the average number of iterations needed to find, for the first time, at least 75% of the inliers and it is averaged over 1000 runs per frame on a sample sequence. Whereas in [11] only w 2 is used, we prove that w 1 is a stronger cue and their combination exceeds both. The best performance is obtained by far with the complete guided sampling, reducing the number of iterations by up to ten times as the inlier ratio decreases.…”

“…Therefore, object detection fails in case of wide baseline matching. Secondly, by updating the model description the model size remains constant, and it is more efficient than the brute force approach of adding features recovered from synthetic views [12,11]. In the latter, the increase in size and the many wrong matches generated by synthetic views having similar appearance, respectively, need additional countermeasures.…”

“…Since features are not perspectively invariant, object recognition fails in case of wide baseline matching. Many approaches handle this by adding synthetic features the training images [12,11]. However, this makes feature matching ambiguous as the number of features grows.…”

“…Recent approaches based on this paradigm rely on feature discriminability for correct matches and on the robust estimator capabilities for outlier rejection [2,11,12]. However, this presents numerous critical issues: 1.…”

“…Robust estimators, like RANSAC [8], generate pose hypotheses without exploiting any information contained in the model, thus making hypothesis generation prone to the potential inconsistency of the matches. In [11], a simple guided sampling based on co-visibility among correspondences is used, without investigating on other cues. In [14], priority is used in the matching stage by considering only the most recurring features.…”