Academic
Publications
Online learning of patch perspective rectification for efficient object detection

Online learning of patch perspective rectification for efficient object detection,10.1109/CVPR.2008.4587514,Stefan Hinterstoisser,Selim Benhimane,Nass

Online learning of patch perspective rectification for efficient object detection   (Citations: 12)
BibTex | RIS | RefWorks Download
For a large class of applications, there is time to train the system. In this paper, we propose a learning-based ap- proach to patch perspective rectification, and show that it is both faster and more reliable than state-of-the-art ad hoc affine region detection methods. Our method performs in three steps. First, a classifier provides for every keypoint not only its identity, but also a first estimate of its transformation. This estimate allows carrying out, in the second step, an accurate perspective rectification using linear predictors. We show that both the classifier and the linear predictors can be trained online, which makes the approach convenient. The last step is a fast verification -made possible by the accurate perspective rectification- of the patch identity and its sub-pixel preci- sion position estimation. We test our approach on real-time 3D object detection and tracking applications. We show that we can use the estimated perspective rectifications to determine the object pose and as a result, we need much fewer correspondences to obtain a precise pose estimation.
Cumulative Annual
View Publication
The following links allow you to view full publications. These links are maintained by other sources not affiliated with Microsoft Academic Search.
    • ...The first method was called LEOPAR in Hinterstoisser et al. (2008) where it was originally published, and will be refer as ALGO1 in this paper...
    • ...ALGO1 Hinterstoisser et al. (2008) 1.05 seconds...

    Stefan Hinterstoisseret al. Learning Real-Time Perspective Patch Rectification

    • ...DOT (for Dominant Orientation Templates) to Affine Region Detectors [12] (Harris-Affine, Hessian-Affine, MSER, IBR, EBR), to patch rectification methods [8, 7, 6] (Leopar, Panter, Gepard) and to the Histograms-of-Gradients (HoG) template matching approach [1]...
    • ...As it was done in [7], in Figure 4(c), we compare the average overlap between the ground truth quadrangles and their corresponding warped versions obtained with DOT, HoG, the patch rectification methods and with the affine regions detectors...
    • ...Since the Affine Region Detectors deliver elliptic regions we fit quadrangles around these ellipses by aligning them to the main gradient orientation as it was done in [7]...
    • ...Due to the robustness and the real-time capability of our approach, DOT is suited for many different applications including untextured object detection as shown in Figure 8, and planar patches detection as shown in Figure 9. Although neither a final refinement nor any final verification, by contrast with [7] for example, was applied to the found 3D objects, the results are very accurate, robust and stable...

    Stefan Hinterstoisseret al. Dominant orientation templates for real-time detection of texture-less...

    • ...[4] proposed a method with a secondary classifier to obtain the perspective transformation of a signal patch...

    Ce Gaoet al. Multilayer Ferns: A Learning-based Approach of Patch Recognition and H...

    • ...In order to remove this restriction we use the approach presented by Hinterstoisser et al. [11], which allows to update the matrix SI in a way such that we can increase the number of random transformations nt without the necessity to recompute the updated ˆI from scratch...

    Stefan Holzeret al. Adaptive Linear Predictors for Real-Time Tracking

    • ...Recent work [5, 6] showed that learning-based patch rectification methods are both faster and more reliable than affine region methods...
    • ...More recently, an approach based on learning instead of ad hoc detectors was developed by Hinterstoisser et al. [6, 5]. This approach appears to be more reliable and much faster, but relies on an extensive training stage...
    • ...The approach of [6, 5] is made of two steps...
    • ...For the first step, [6] uses the Ferns classifier [11] while [5] relies on linear classifiers...
    • ...Our second step is similar to the one in [6] and [5] but we show that the precomputation of the sample points for each training sample makes real-time computation of the linear regressors possible...
    • ...The method proposed in [6] performs the other way around, in two steps...
    • ...Contrary to [5, 6], we do not want to learn the patches in a long computational expensive offline training phase, but online and in real-time without loosing the ability to accurately estimate the pose of the patches under large viewpoint changes...
    • ...The refinement of the keypoint pose is performed very similarly to [5, 6] with linear predictors but we use a simple trick to speed up the predictors training...
    • ...This step is similar to what was done in [5, 6]. We describe it quickly for completeness but more details can be found in these papers...
    • ...Finally, as it was done in [5, 6], we select the correct match consisting of the retrieved pose and the feature point identity based on the cross-correlation between the normalized reference patch n(pi) and the normalized warped one after refinement...
    • ...Note these methods have already been proved superior to affine region methods [6, 5]. Top graph: Correct identity and coarse pose estimation percentage against viewpoint change on the Graffiti image set...
    • ...We limited the comparison of our approach to the other learning-based methods called Leopar and Panter because these methods have already been proved superior to affine region methods in the related papers [6, 5]. We did this comparison on many synthetically warped Graffiti images [9] to obtain a statistically significant statement...
    • ...The results are shown in the three graphs of Fig. 5 and Fig. 4. Our performance measure for the graph in Fig. 4 and for the top Leopar [6] 1.05 seconds1...
    • ...Gepard performs similarly to Panter [5] and better than Leopar [6]...

    Stefan Hinterstoisseret al. Real-time learning of accurate patch rectification

Sort by: