Unsupervised Domain Adaptation (UDA) is essential for enabling semantic segmentation in new domains without requiring costly pixel-wise annotations. State-of-the-art (SOTA) UDA methods primarily use self-training with architecturally identical teacher and student networks, relying on Exponential Moving Average (EMA) updates. However, these approaches face substantial performance degradation with lightweight models due to inherent architectural inflexibility leading to low-quality pseudo-labels. To address this, we propose Distilled Unsupervised Domain Adaptation (DUDA), a novel framework that combines EMA-based self-training with knowledge distillation (KD). Our method employs an auxiliary student network to bridge the architectural gap between heavyweight and lightweight models for EMA-based updates, resulting in improved pseudo-label quality. DUDA employs a strategic fusion of UDA and KD, incorporating innovative elements such as gradual distillation from large to small networks, inconsistency loss prioritizing poorly adapted classes, and learning with multiple teachers. Extensive experiments across four UDA benchmarks demonstrate DUDA's superiority in achieving SOTA performance with lightweight models, often surpassing the performance of heavyweight models from other approaches.

We propose Graph2Nav, a real-time 3D object-relation graph generation framework, for autonomous navigation in the real world. Our framework fully generates and exploits both 3D objects and a rich set of semantic relationships among objects in a 3D layered scene graph, which is applicable to both indoor and outdoor scenes. It learns to generate 3D semantic relations among objects, by leveraging and advancing state-of-the-art 2D panoptic scene graph works into the 3D world via 3D semantic mapping techniques. This approach avoids previous training data constraints in learning 3D scene graphs directly from 3D data. We conduct experiments to validate the accuracy in locating 3D objects and labeling object-relations in our 3D scene graphs. We also evaluate the impact of Graph2Nav via integration with SayNav, a state-of-the-art planner based on large language models, on an unmanned ground robot to object search tasks in real environments. Our results demonstrate that modeling object relations in our scene graphs improves search efficiency in these navigation tasks.

Deep learning-based solutions for semantic segmentation suffer from significant performance degradation when tested on data with different characteristics than what was used during the training. Adapting the models using annotated data from the new domain is not always practical. Unsupervised Domain Adaptation (UDA) approaches are crucial in deploying these models in the actual operating conditions. Recent state-of-the-art (SOTA) UDA methods employ a teacher-student self-training approach, where a teacher model is used to generate pseudo-labels for the new data which in turn guide the training process of the student model. Though this approach has seen a lot of success, it suffers from the issue of noisy pseudo-labels being propagated in the training process. To address this issue, we propose an auxiliary pseudo-label refinement network (PRN) for online refining of the pseudo labels and also localizing the pixels whose predicted labels are likely to be noisy. Being able to improve the quality of pseudo labels and select highly reliable ones, PRN helps self-training of segmentation models to be robust against pseudo label noise propagation during different stages of adaptation. We evaluate our approach on benchmark datasets with three different domain shifts, and our approach consistently performs significantly better than the previous state-of-the-art methods.

Unsupervised domain adaptation (UDA) approaches focus on adapting models trained on a labeled source domain to an unlabeled target domain. In contrast to UDA, source-free domain adaptation (SFDA) is a more practical setup as access to source data is no longer required during adaptation. Recent state-of-the-art (SOTA) methods on SFDA mostly focus on pseudo-label refinement based self-training which generally suffers from two issues: i) inevitable occurrence of noisy pseudo-labels that could lead to early training time memorization, ii) refinement process requires maintaining a memory bank which creates a significant burden in resource constraint scenarios. To address these concerns, we propose C-SFDA, a curriculum learning aided self-training framework for SFDA that adapts efficiently and reliably to changes across domains based on selective pseudo-labeling. Specifically, we employ a curriculum learning scheme to promote learning from a restricted amount of pseudo labels selected based on their reliabilities. This simple yet effective step successfully prevents label noise propagation during different stages of adaptation and eliminates the need for costly memory-bank based label refinement. Our extensive experimental evaluations on both image recognition and semantic segmentation tasks confirm the effectiveness of our method. C-SFDA is also applicable to online test-time domain adaptation and outperforms previous SOTA methods in this task.

Precise estimation of global orientation and location is critical to ensure a compelling outdoor Augmented Reality (AR) experience. We address the problem of geo-pose estimation by cross-view matching of query ground images to a geo-referenced aerial satellite image database. Recently, neural network-based methods have shown state-of-the-art performance in cross-view matching. However, most of the prior works focus only on location estimation, ignoring orientation, which cannot meet the requirements in outdoor AR applications. We propose a new transformer neural network-based model and a modified triplet ranking loss for joint location and orientation estimation. Experiments on several benchmark cross-view geo-localization datasets show that our model achieves state-of-the-art performance. Furthermore, we present an approach to extend the single image query-based geo-localization approach by utilizing temporal information from a navigation pipeline for robust continuous geo-localization. Experimentation on several large-scale real-world video sequences demonstrates that our approach enables high-precision and stable AR insertion.

This paper presents a novel approach for the Vision-and-Language Navigation (VLN) task in continuous 3D environments, which requires an autonomous agent to follow natural language instructions in unseen environments. Existing end-to-end learning-based VLN methods struggle at this task as they focus mostly on utilizing raw visual observations and lack the semantic spatio-temporal reasoning capabilities which is crucial in generalizing to new environments. In this regard, we present a hybrid transformer-recurrence model which focuses on combining classical semantic mapping techniques with a learning-based method. Our method creates a temporal semantic memory by building a top-down local ego-centric semantic map and performs cross-modal grounding to align map and language modalities to enable effective learning of VLN policy. Empirical results in a photo-realistic long-horizon simulation environment show that the proposed approach outperforms a variety of state-of-the-art methods and baselines with over 22% relative improvement in SPL in prior unseen environments.

Understanding the geometric relationships between objects in a scene is a core capability in enabling both humans and autonomous agents to navigate in new environments. A sparse, unified representation of the scene topology will allow agents to act efficiently to move through their environment, communicate the environment state with others, and utilize the representation for diverse downstream tasks. To this end, we propose a method to train an autonomous agent to learn to accumulate a 3D scene graph representation of its environment by simultaneously learning to navigate through said environment. We demonstrate that our approach, GraphMapper, enables the learning of effective navigation policies through fewer interactions with the environment than vision-based systems alone. Further, we show that GraphMapper can act as a modular scene encoder to operate alongside existing Learning-based solutions to not only increase navigational efficiency but also generate intermediate scene representations that are useful for other future tasks.

Class imbalance is a fundamental problem in computer vision applications such as semantic segmentation. Specifically, uneven class distributions in a training dataset often result in unsatisfactory performance on under-represented classes. Many works have proposed to weight the standard cross entropy loss function with pre-computed weights based on class statistics, such as the number of samples and class margins. There are two major drawbacks to these methods: 1) constantly up-weighting minority classes can introduce excessive false positives in semantic segmentation; 2) a minority class is not necessarily a hard class. The consequence is low precision due to excessive false positives. In this regard, we propose a hard-class mining loss by reshaping the vanilla cross entropy loss such that it weights the loss for each class dynamically based on instantaneous recall performance. We show that the novel recall loss changes gradually between the standard cross entropy loss and the inverse frequency weighted loss. Recall loss also leads to improved mean accuracy while offering competitive mean Intersection over Union (IoU) performance. On Synthia dataset11Synthia-Sequence Summer split, recall loss achieves 9% relative improvement on mean accuracy with competitive mean IoU using DeepLab-ResNet18 compared to the cross entropy loss. Code available at https://github.com/PotatoTian/recall-semseg.

Understanding the perceived scene during navigation enables intelligent robot behaviors. Current vision-based semantic SLAM (Simultaneous Localization and Mapping) systems provide these capabilities. However, their performance decreases in visually-degraded environments, that are common places for critical robotic applications, such as search and rescue missions. In this paper, we present SIGNAV, a real-time semantic SLAM system to operate in perceptually-challenging situations. To improve the robustness for navigation in dark environments, SIGNAV leverages a multi-sensor navigation architecture to fuse vision with additional sensing modalities, including an inertial measurement unit (IMU), LiDAR, and wheel odometry. A new 2.5D semantic segmentation method is also developed to combine both images and LiDAR depth maps to generate semantic labels of 3D mapped points in real time. We demonstrate that the navigation accuracy from SIGNAV in a variety of indoor environments under both normal lighting and dark conditions. SIGNAV also provides semantic scene understanding capabilities in visually-degraded environments. We also show the benefits of semantic information to SIGNAV’s performance.

Visual navigation for autonomous agents is a core task in the fields of computer vision and robotics. Learning-based methods, such as deep reinforcement learning, have the potential to outperform the classical solutions developed for this task; however, they come at a significantly increased computational load. Through this work, we design a novel approach that focuses on performing better or comparable to the existing learning-based solutions but under a clear time/computational budget. To this end, we propose a method to encode vital scene semantics such as traversable paths, unexplored areas, and observed scene objects–alongside raw visual streams such as RGB, depth, and semantic segmentation masks—into a semantically informed, top-down egocentric map representation. Further, to enable the effective use of this information, we introduce a novel 2-D map attention mechanism, based on the successful multi-layer Transformer networks. We conduct experiments on 3-D reconstructed indoor PointGoal visual navigation and demonstrate the effectiveness of our approach. We show that by using our novel attention schema and auxiliary rewards to better utilize scene semantics, we outperform multiple baselines trained with only raw inputs or implicit semantic information while operating with an 80% decrease in the agent’s experience.

Prior works on text-based video moment localization focus on temporally grounding the textual query in an untrimmed video. These works assume that the relevant video is already known and attempt to localize the moment on that relevant video only. Different from such works, we relax this assumption and address the task of localizing moments in a corpus of videos for a given sentence query. This task poses a unique challenge as the system is required to perform: 2) retrieval of the relevant video where only a segment of the video corresponds with the queried sentence, 2) temporal localization of moment in the relevant video based on sentence query. Towards overcoming this challenge, we propose Hierarchical Moment Alignment Network (HMAN) which learns an effective joint embedding space for moments and sentences. In addition to learning subtle differences between intra-video moments, HMAN focuses on distinguishing inter-video global semantic concepts based on sentence queries. Qualitative and quantitative results on three benchmark text-based video moment retrieval datasets - Charades-STA, DiDeMo, and ActivityNet Captions - demonstrate that our method achieves promising performance on the proposed task of temporal localization of moments in a corpus of videos.

Proper occlusion based rendering is very important to achieve realism in all indoor and outdoor Augmented Reality (AR) applications. This paper addresses the problem of fast and accurate dynamic occlusion reasoning by real objects in the scene for large scale outdoor AR applications. Conceptually, proper occlusion reasoning requires an estimate of depth for every point in augmented scene which is technically hard to achieve for outdoor scenarios, especially in the presence of moving objects. We propose a method to detect and automatically infer the depth for real objects in the scene without explicit detailed scene modeling and depth sensing (e.g. without using sensors such as 3D-LiDAR). Specifically, we employ instance segmentation of color image data to detect real dynamic objects in the scene and use either a top-down terrain elevation model or deep learning based monocular depth estimation model to infer their metric distance from the camera for proper occlusion reasoning in real time. The realized solution is implemented in a low latency real-time framework for video-see-though AR and is directly extendable to optical-see-through AR. We minimize latency in depth reasoning and occlusion rendering by doing semantic object tracking and prediction in video frames.

In this paper, we address the problem of utilizing web images in training robust joint embedding models for the image-text retrieval task. Prior webly supervised approaches directly leverage weakly annotated web images in the joint embedding learning framework. The objective of these approaches would suffer significantly when the ratio of noisy and missing tags associated with the web images is very high. In this regard, we propose a CP decomposition based tensor completion framework to refine the tags of web images by modeling observed ternary inter-relations between the sets of labeled images, tags, and web images as a tensor. To effectively deal with the high ratio of missing entries likely in our case, we incorporate intra-modal correlation as side information in the proposed framework. Our tag refinement approach combined with existing web supervised image-text embedding approaches provide a more principled way for learning the joint embedding models in the presence of significant noise from web data and limited clean labeled data. Experiments on benchmark datasets demonstrate that the proposed approach helps to achieve a significant performance gain in image-text retrieval.

We study an important, yet largely unexplored problem of large-scale cross-modal visual localization by matching ground RGB images to a geo-referenced aerial LIDAR 3D point cloud (rendered as depth images). Prior works were demonstrated on small datasets and did not lend themselves to scaling up for large-scale applications. To enable large-scale evaluation, we introduce a new dataset containing over 550K pairs (covering 143 km2 area) of RGB and aerial LIDAR depth images. We propose a novel joint embedding based method that effectively combines the appearance and semantic cues from both modalities to handle drastic cross-modal variations. Experiments on the proposed dataset show that our model achieves a strong result of a median rank of 5 in matching across a large test set of 50K location pairs collected from a 14km^2 area. This represents a significant advancement over prior works in performance and scale. We conclude with qualitative results to highlight the challenging nature of this task and the benefits of the proposed model. Our work provides a foundation for further research in cross-modal visual localization.

Image datasets play a pivotal role in advancing computer vision and multimedia research. However, most of the datasets are created by extensive human effort and are extremely expensive to scale-up. To address these issues, several automatic and semi-automatic approaches have been proposed for creating datasets by refining web images. However, these approaches either include significant redundant images in the dataset or fail to provide a diverse enough set to train a robust classifier. Ideally, a representative subset should be both semantically and visually diverse so as to provide the maximum amount of information under the current budget. Most current approaches are entirely based on the analysis of visual features, which may not correlate well with image semantics, and hence, collected images may not be sufficient to give a detailed understanding of a category. In this paper, we propose a system for creating diverse image dataset collections from the web with limited manual labeling effort. It is based upon a semi-supervised sparse coding framework that employs a joint visual-semantic space to simultaneously utilize both the images and associated textual information from the web for dataset construction. In addition, the proposed system is online and is capable of collecting more discriminative images continuously as new data becomes available, which is also suitable for enriching the existing datasets. The experiments demonstrate that our system can create and enrich datasets with limited manual labeling, with better cross-dataset generalization capability and diversity compared to the state-of-the-art datasets.

There have been a few recent methods proposed in text to video moment retrieval using natural language queries, but requiring full supervision during training. However, acquiring a large number of training videos with temporal boundary annotations for each text description is extremely time-consuming and often not scalable. In order to cope with this issue, in this work, we introduce the problem of learning from weak labels for the task of text to video moment retrieval. The weak nature of the supervision is because, during training, we only have access to the video-text pairs rather than the temporal extent of the video to which different text descriptions relate. We propose a joint visual-semantic embedding based framework that learns the notion of relevant segments from video using only video-level sentence descriptions. Specifically, our main idea is to utilize latent alignment between video frames and sentence descriptions using Text-Guided Attention (TGA). TGA is then used during the test phase to retrieve relevant moments. Experiments on two benchmark datasets demonstrate that our method achieves comparable performance to state-of-the-art fully supervised approaches.

The identity of subjects in many portraits has been a matter of debate for art historians that relied upon subjective analysis of facial features to resolve ambiguity in sitter identity. Developing automated face verification technique has thus garnered interest to provide a quantitative way to reinforce the decision arrived at by the art historians. However, most existing works often fail to resolve ambiguities concerning the identity of the subjects due to significant variation in artistic styles and the limited availability and authenticity of art images. To these ends, we explore the use of deep Siamese Convolutional Neural Networks (CNN) to provide a measure of similarity between a pair of portraits. To mitigate limited training data issue, we employ CNN based style-transfer technique that creates several new images by recasting an art style to other images, keeping original image content unchanged. The resulting system thereby learns features which are discriminative and invariant to changes in artistic styles. Our approach shows significant improvement over baselines and state-of-the-art methods on several examples which are identified by art historians as being very challenging and controversial.

Detecting objects that are carried when someone enters or exits a room is very useful for a wide range of smart building applications including safety, security, and energy efficiency. While there has been a significant amount of work on object recognition using large-scale RGB image datasets, RGB cameras are too privacy invasive in many smart building applications and they work poorly in the dark. Additionally, deep object detection networks require powerful and expensive GPUs. We propose a novel system that we call ODDS (Object Detector using a Depth Sensor) that can detect objects in real-time using only raw depth data on an embedded GPU, e.g., NVIDIA Jetson TX1. Hence, our solution is significantly less privacy invasive (even if the sensor is compromised) and less expensive, while maintaining a comparable accuracy with state of the art solutions. Specifically, we resort to training a deep convolutional neural network using raw depth images, with curriculum based learning to improve accuracy by considering the complexity and imbalance in object classes and developing a sparse coding based technique that speeds up the system ~2× with minimal loss of accuracy. Based on a complete implementation and real-world evaluation, we see ODDS achieve 80.14% mean average precision in object detection in real-time (5-6 FPS) on a Jetson TX1.

Constructing a joint representation invariant across different modalities (e.g., video, language) is of significant importance in many multimedia applications. While there are a number of recent successes in developing effective image-text retrieval methods by learning joint representations, the video-text retrieval task, however, has not been explored to its fullest extent. In this paper, we study how to effectively utilize available multimodal cues from videos for the cross-modal video-text retrieval task. Based on our analysis, we propose a novel framework that simultaneously utilizes multi-modal features (different visual characteristics, audio inputs, and text) by a fusion strategy for efficient retrieval. Furthermore, we explore several loss functions in training the embedding and propose a modified pairwise ranking loss for the task. Experiments on MSVD and MSR-VTT datasets demonstrate that our method achieves significant performance gain compared to the state-of-the-art approaches.

Cross-modal retrieval between visual data and natural language description remains a long-standing challenge in multimedia. While recent image-text retrieval methods offer great promise by learning deep representations aligned across modalities, most of these methods are plagued by the issue of training with small-scale datasets covering a limited number of images with ground-truth sentences. Moreover, it is extremely expensive to create a larger dataset by annotating millions of images with sentences and may lead to a biased model. Inspired by the recent success of webly supervised learning in deep neural networks, we capitalize on readily-available web images with noisy annotations to learn robust image-text joint representation. Specifically, our main idea is to leverage web images and corresponding tags, along with fully annotated datasets, in training for learning the visual-semantic joint embedding. We propose a two-stage approach for the task that can augment a typical supervised pair-wise ranking loss based formulation with weakly-annotated web images to learn a more robust visual-semantic embedding. Experiments on two standard benchmark datasets demonstrate that our method achieves a significant performance gain in image-text retrieval compared to state-of-the-art approaches.

Constructing a feature representation invariant to certain types of geometric and photometric transformations is of significant importance in many computer vision applications. In spite of significant effort, developing invariant feature representations remains a challenging problem. Most of the existing representations often fail to satisfy the longterm repeatability requirements of specific applications like vision-based localization, applications whose domain includes significant, non-uniform illumination and environmental changes. To these ends, we explore the use of natural image pairs (i.e. images captured of the same location but at different times) as an additional source of supervision to generate an improved feature representation for the task of vision-based localization. Specifically, we resort to training deep denoising autoencoder, with CNN feature representation of one image in the pair being treated as a noisy version of the other. The resulting system thereby learns localization features which are both discriminative and invariant to illumination and environmental changes. In experiments tailored towards vision-based localization, features generated using the proposed method produced higher matching rates than state-of-the-art image features.

Most video summarization approaches have focused on extracting a summary from a single video; we propose an unsupervised framework for summarizing a collection of videos. We observe that each video in the collection may contain some information that other videos do not have, and thus exploring the underlying complementarity could be beneficial in creating a diverse informative summary. We develop a novel diversity-aware sparse optimization method for multi-video summarization by exploring the complementarity within the videos. Our approach extracts a multi-video summary, which is both interesting and representative in describing the whole video collection. To efficiently solve our optimization problem, we develop an alternating minimization algorithm that minimizes the overall objective function with respect to one video at a time while fixing the other videos. Moreover, we introduce a new benchmark data set, Tour20, that contains 140 videos with multiple manually created summaries, which were acquired in a controlled experiment. Finally, by extensive experiments on the new Tour20 data set and several other multi-view data sets, we show that the proposed approach clearly outperforms the state-of-the-art methods on the two problems-topic-oriented video summarization and multi-view video summarization in a camera network.

This paper presents the Online Social Network Investigator (OSNI), a scalable distributed system to search social net- work data, based on a spatiotemporal window and a list of keywords. Given that only 2% of tweets are geolocated, we have implemented and compared various state-of-art loca- tion estimation techniques. Further, to enrich the context of posts, associations of images to terms are estimated through various classication techniques. The accuracies of these es- timations are evaluated on large real datasets. OSNI's query interface is available on the Web.

Image datasets play a pivotal role in advancing multimedia and image analysis research. However, most of these datasets are created by extensive human effort and extremely expensive to scale up. There is high chance that we may have no instances for some required concepts in these data-sets or the available instances do not cover the diversity of real-world scenarios. In this regard, several approaches for learning from web images and refining them have been proposed, but these approaches either include significant redundant instances in the dataset or fail to guarantee a diverse enough set to train a robust classifier. In this work, we propose a semi-supervised sparse coding framework to collect a diverse set of images with minimal human effort, which can be used to both create a dataset from scratch or enrich an existing dataset with diverse examples. To evaluate our method, we constructed an image dataset with our framework, which is named as DivNet. Experiments on this dataset demonstrate that our method not only reduces manual effort, but also the created dataset has excellent accuracy, diversity and cross-dataset generalization ability.

Detection and classification of vehicles are two of the most challenging tasks of a video-based intelligent transportation system. Traditional detection and classification methods are computationally highly expensive and become unsuccessful in many cases such as occlusion among the vehicles and when differences between pixel intensities of vehicles and backgrounds are small. In this paper, a novel detection and classification method is proposed using multiple time-spatial images (TSIs), each obtained from a virtual detection line on the frames of a video. Such a use of multiple TSIs provides the opportunity to identify the latent occlusions among the vehicles and to reduce the dependencies of the pixel intensities between the still and moving objects to increase the accuracy of detection performance as well as to achieve an improved classification performance. In order to identify the class of a particular vehicle, a two-step k nearest neighborhood classification scheme is proposed by utilizing the shape-based, shape-invariant, and texture-based features of the segmented regions corresponding to the vehicle appeared in appropriate frames that are determined from the TSIs of the video. Extensive experimentations are carried out in vehicular traffics of varying environments to evaluate the detection and classification performance of the proposed method, as compared with the existing methods. Experimental results demonstrate that the proposed method provides a significant improvement in counting and classifying the vehicles in terms of accuracy and robustness alongside a substantial reduction of execution time, as compared with that of the other methods.