favorite2Strong Geometric Context for Detection Estimating camera pose of a test image with respect to a 3D GIS model aids in reasoning about the geometric validity of hypothesized object detections (e.g. pruning false positives in unlikely locations or boosting low-scoring candidates in likely regions).
favorite4Compared to these works, our pipeline utilizes full 6D camera pose estimation and a richer scene model derived from the GIS map that supports not only improved detection rescoring but also depth and semantic label priors.
favorite7Secondly, we show that simple cues derived from the GIS model can provide significantly improved performance over baselines for depth estimation, pedestrian detection and semantic segmentation.
favorite27First, we demonstrate a pipeline that performs precise resectioning of test images against GIS map data to generate geosemantic context applicable to multiple scene understanding tasks.
favorite24This fused geocontext provides a basis for efficiently transferring rich geometric and semantic information to a novel test image where it is used to improve performance of general scene understanding (depth, detection, and segmentation)..
favorite13We then demonstrate the utility of these contextual constraints for re-scoring pedestrian detections, and use these GIS contextual features alongside object detection score maps to improve a CRF-based semantic segmentation framework, boosting accuracy over baseline models..
favorite6We achieve perfect location recognition results on the Dubrovnik dataset using a random subset of 200 query features that pass the k-ratio and cluster-wise ratio tests, suggesting that our approach successfully finds local discriminative correspondences for all 800 test images.
favorite13For any such candidate matching view, we search for the first and second nearest neighbor matches using a kd-tree built over the query image features and apply the 1-ratio test.
favorite7In this paper we consider the problem of estimating the full 6DOF camera pose of a query image with respect to a large-scale 3D model such as those obtained from a Structure-from-Motion (SfM) pipeline [27, 34, 16, 25].
favorite0With a smaller model and prioritized search, it becomes possible to replace the traditional approach of 2D-to-3D forward matching, with 3D-to-2D back matching, allowing the ratio test to be performed in the sparser feature space of the query image.
favorite5Standard heuristics for identifying distinctive matches, such as the distance ratio-test of Lowe , which compares the distance to the nearestneighbor point descriptor with that of the second-nearest neighbor, fail due to proximity of other model feature descriptors.