Visual localization is a fundamental machine learning problem. Absolute Pose Regression (APR) trains a scene-dependent model to efficiently map an input image to the camera pose in a pre-defined scene. However, many applications have continually changing environments, where inference data at novel poses or scene conditions (weather, geometry) appear after deployment. Training APR on a fixed dataset leads to overfitting, making it fail catastrophically on challenging novel data. This work proposes Continual Domain Expansion (ConDo), which continually collects unlabeled inference data to update the deployed APR. Instead of applying standard unsupervised domain adaptation methods which are ineffective for APR, ConDo effectively learns from unlabeled data by distilling knowledge from scene-agnostic localization methods. By sampling data uniformly from historical and newly collected data, ConDo can effectively expand the generalization domain of APR. Large-scale benchmarks with various scene types are constructed to evaluate models under practical (long-term) data changes. ConDo consistently and significantly outperforms baselines across architectures, scene types, and data changes. On challenging scenes (Fig.1), it reduces the localization error by >7x (14.8m vs 1.7m). Analysis shows the robustness of ConDo against compute budgets, replay buffer sizes and teacher prediction noise. Comparing to model re-training, ConDo achieves similar performance up to 25x faster.

Image matching is a fundamental computer vision problem. While learning-based methods achieve state-of-the-art performance on existing benchmarks, they generalize poorly to in-the-wild images. Such methods typically need to train separate models for different scene types (e.g., indoor vs. outdoor) and are impractical when the scene type is unknown in advance. One of the underlying problems is the limited scalability of existing data construction pipelines, which limits the diversity of standard image matching datasets. To address this problem, we propose GIM, a self-training framework for learning a single generalizable model based on any image matching architecture using internet videos, an abundant and diverse data source. Given an architecture, GIM first trains it on standard domain-specific datasets and then combines it with complementary matching methods to create dense labels on nearby frames of novel videos. These labels are filtered by robust fitting, and then enhanced by propagating them to distant frames. The final model is trained on propagated data with strong augmentations. Not relying on complex 3D reconstruction makes GIM much more efficient and less likely to fail than standard SfM-and-MVS based frameworks. We also propose ZEB, the first zero-shot evaluation benchmark for image matching. By mixing data from diverse domains, ZEB can thoroughly assess the cross-domain generalization performance of different methods. Experiments demonstrate the effectiveness and generality of GIM. Applying GIM consistently improves the zero-shot performance of 3 state-of-the-art image matching architectures as the number of downloaded videos increases (Fig. 1 (a)); with 50 hours of YouTube videos, the relative zero-shot performance improves by 6.9% − 18.1%. GIM also enables generalization to extreme cross-domain data such as Bird Eye View (BEV) images of projected 3D point clouds (Fig. 1 (c)). More importantly, our single zero-shot model consistently outperforms domain-specific baselines when evaluated on downstream tasks inherent to their respective domains. The code will be released upon acceptance.

This paper presents a matching network to establish point correspondence between images. We propose a Multi-Arm Network (MAN) capable of learning region overlap and depth, which can greatly improve keypoint matching robustness while bringing an extra 50% of computational time during the inference stage. By adopting a different design from the state-of-the-art learning based pipeline SuperGlue framework, which requires retraining when a different keypoint detector is adopted, our network can directly work with different keypoint detectors without time-consuming retraining processes. Comprehensive experiments conducted on four public benchmarks involving both outdoor and indoor scenarios demonstrate that our proposed MAN outperforms state-of-the-art methods.

This paper proposes a new end-to-end trainable matching network based on receptive field, RF-Net, to compute sparse correspondence between images. Building end-to-end trainable matching framework is desirable and challenging. The very recent approach, LF-Net, successfully embeds the entire feature extraction pipeline into a jointly trainable pipeline, and produces the state-of-the-art matching results. This paper introduces two modifications to the structure of LF-Net. First, we propose to construct receptive feature maps, which lead to more effective keypoint detection. Second, we introduce a general loss function term, neighbor mask, to facilitate training patch selection. This results in improved stability in descriptor training. We trained RF-Net on the open dataset HPatches, and compared it with other methods on multiple benchmark datasets. Experiments show that RF-Net outperforms existing state-of-the-art methods.

This paper presents a novel virtual-real registration approach for augmented reality (AR) in large-scale outdoor environments. Essentially, it is a pose estimation for the mobile camera images (ground camera images) in 3D model recovered by Unmanned Aerial Vehicle (UAV) image sequence via Structure-From-Motion (SFM) technology. The approach considers to indirectly establish the spatial relationship between 2D and 3D space by inferring the transformation relationship between the ground camera images and the UAV 3D model rendered images. Specifically, the proposed approach can overcome the positioning errors, which are deterioration and drift in the GPS, and deviation of orientation. The experimental results demonstrate the possibility of the proposed virtual-real registration approach, and show that the approach is robust, efficient and intuitive for AR in large-scale outdoor environments.

It is challenging to direct extract the feature descriptors of the object in the point cloud, although deep learning has been widely used with the classification and detection in the point cloud, those methods hidden feature presentation in the network. Since the point cloud scanned by the Laser Scanner usually have different point density, unordered and even the different occlusion, which go beyond the reach of handcrafted descriptors, e.g. FPH, FPFH, VFH, ROPS. In this paper, we aim to direct extract the feature descriptors of the point cloud object through the raw point cloud. Inspired by the recent success of the Siamese networks[6], PointNet[7] and PointNet++[8], we propose a novel network to direct extract the feature descriptors of the whole point cloud object. We train our network with the Euclidean distance as the loss function which reflects feature descriptors similarity. The experiment object datasets were acquired by Mobile Laser Scanning (MLS) system which contains 6 categories. Experiment result shows that our network has a robust generalization, which can well direct extract the feature descriptors of the whole point cloud object.

Describing the same scene with different imaging style or rendering image from its 3D model gives us different domain images. Different domain images tend to have a gap and different local appearances, which raise the main challenge on the cross-domain image patch matching. In this paper, we propose to incorporate AutoEncoder into the Siamese network, named as H-Net, of which the structural shape resembles the letter H. The H-Net achieves state-of-the-art performance on the cross-domain image patch matching. Furthermore, we improved H-Net to H-Net++. The H-Net++ extracts invariant feature descriptors in cross-domain image patches and achieves state-of-the-art performance by feature retrieval in Euclidean space. As there is no benchmark dataset including cross-domain images, we made a cross-domain image dataset which consists of camera images, rendering images from UAV 3D model, and images generated by CycleGAN algorithm. Experiments show that the proposed H-Net and H-Net++ outperform the existing algorithms. Our code and cross-domain image dataset are available at https://github.com/Xylon-Sean/H-Net.