Two-stage Learning-to-Defer (L2D) enables optimal task delegation by assigning each input to either a fixed main model or one of several offline experts, supporting reliable decision-making in complex, multi-agent environments. However, existing L2D frameworks assume clean inputs and are vulnerable to adversarial perturbations that can manipulate query allocation—causing costly misrouting or expert overload. We present the first comprehensive study of adversarial robustness in two-stage L2D systems. We introduce two novel attack strategies—untargeted and targeted—which respectively disrupt optimal allocations or force queries to specific agents. To defend against such threats, we propose SARD, a convex learning algorithm built on a family of surrogate losses that are provably Bayes-consistent and $(\mathcal{R}, \mathcal{G})$-consistent. These guarantees hold across classification, regression, and multi-task settings. Empirical results demonstrate that SARD significantly improves robustness under adversarial attacks while maintaining strong clean performance, marking a critical step toward secure and trustworthy L2D deployment.

The Two-Stage Learning-to-Defer (L2D) framework has been extensively studied for classification and, more recently, regression tasks. However, many real-world applications require solving both tasks jointly in a multi-task setting. We introduce a novel Two-Stage L2D framework for multi-task learning that integrates classification and regression through a unified deferral mechanism. Our method leverages a two-stage surrogate loss family, which we prove to be both Bayes-consistent and $(\mathcal{G}, \mathcal{R})$-consistent, ensuring convergence to the Bayes-optimal rejector. We derive explicit consistency bounds tied to the cross-entropy surrogate and the $L_1$-norm of agent-specific costs, and extend minimizability gap analysis to the multi-expert two-stage regime. We also make explicit how shared representation learning—commonly used in multi-task models—affects these consistency guarantees. Experiments on object detection and electronic health record analysis demonstrate the effectiveness of our approach and highlight the limitations of existing L2D methods in multi-task scenarios.

The growth of online educational content, particularly slide-based video lectures, has created a need for tools that enhance navigation, comprehension, and accessibility. Many existing systems for video analysis are closed-source, hindering reproducibility and extension. To address this, we present the Lecture Video Analysis Toolkit, an open-source, proof-of-concept application designed for the multimodal analysis of slide video lectures. The toolkit integrates a processing pipeline that includes scene detection, visual entity extraction, transcription, optical character recognition (OCR), and semantic linking between spoken and visual content using embeddings. A key contribution is its interactive interface, motivated by early user feedback, that allows for customization of the viewing experience to suit individual preferences. The entire system is openly available and serves as a research prototype for validating the potential of multimodal analysis in creating more inclusive and improved learning experiences. A live demo is accessible at https://travis-seng.fr/svla, and the source code is openly available at https://github.com/travisseng/svla-toolkit.

We propose to detect defects in old movies, as the first step of a larger framework of old movies restoration by inpainting techniques. The specificity of our work is to learn a film restorer's expertise from a pair of sequences, composed of a movie with defects, and the same movie which was semiautomatically restored with the help of a specialized software. In order to detect those defects with minimal human interaction and further reduce the time spent for a restoration, we feed a U-Net with consecutive defective frames as input to detect the unexpected variations of pixel intensity over space and time. Since the output of the network is a mask of defect location, we first have to create the dataset of mask frames on the basis of restored frames from the software used by the film restorer, instead of classical synthetic ground truth, which is not available. These masks are estimated by computing the absolute difference between restored frames and defectuous frames, combined with thresholding and morphological closing. Our network succeeds in automatically detecting real defects with more precision than the manual selection with an all-encompassing shape, including some the expert restorer could have missed for lack of time.

In this paper, we propose an approach for segmentation of challenging depth images. We first use a semi-automatic segmentation algorithm that only takes a user-defined rectangular area as an input. The quality of the segmentation is very heterogeneous at this stage, and unsufficient to efficiently train a neural network. We thus introduce a learning process that takes this imperfect nature of data into account, by iteratively filtering the dataset to only keep the best segmented images. We show this method improves the neural network's performance by a significant amount.

Macrogroove is an interactive and playful multimedia system that allows the user to play a sound coded in the form of a 3D-sonagram. The user is invited to manually move a laser sheet over this "pseudo-relief", in order to playback the original sound in real time.

Navigating in a 3D networked virtual environment with a six-degree of freedom on a mobile device can be disorientating and challenging to users. In this technical demonstration, we show how 3D bookmarks can be used to simplify such interactions. Our system integrates the 3D bookmarks into a DASH-based network virtual environment in a DASH-compliant manner and is available on both the desktop and mobile.

We demonstrate the use of DASH, a widely-deployed standard for streaming video content, for streaming 3D content in an NVE (Networked Virtual Environment) consisting of 3D geometry and associated textures. We have developed a DASH client for NVE to show how NVE benefits from the advantages of DASH: it offers a scalable, easy-to-deploy 3D streaming framework. In our system, the 3D content is first statically partitioned into compliant DASH data, and metadata is provided in order for the client to manage which data to download. Based on a proposed utility metric for geometry and texture at the different resolution, the client can choose the content to request depending on its viewpoint. We effectively provide a Web-based client to navigate through our sample 3D scene, while deriving the streaming requests from its computation of the necessary online parameters, in a receiver-driven manner.

DASH is now a widely deployed standard for streaming video content due to its simplicity, scalability, and ease of deployment. In this paper, we explore the use of DASH for a different type of media content -- networked virtual environment (NVE), with different properties and requirements. We organize a polygon soup with textures into a structure that is compatible with DASH MPD (Media Presentation Description), with a minimal set of view-independent metadata for the client to make intelligent decisions about what data to download at which resolution. We also present a DASH-based NVE client that uses a view-dependent and network dependent utility metric to decide what to download, based only on the information in the MPD file. We show that DASH can be used on NVE for 3D content streaming. Our work opens up the possibility of using DASH for highly interactive applications, beyond its current use in video streaming.

We introduce a new application for annotating images, with the purpose of constituting training datasets for machine learning algorithms. Our open-source software is meant to be easily used and deployed, configured to meet the annotation needs of any use case, and embeddable in crowdsourcing campaigns using the Amazon Mechanical Turk service.

This paper presents a new method to measure a rotor speed using a smartphone camera. An image processing algorithm is proposed to measure the rotation speed, based on the periodic detection of a mark placed on the rotor. Experiments demonstrate our method can estimate rotor speed with a good precision, provided that the camera frame rate is high enough. The method could be very useful for periodic speed measurement in applications where a sensor is prohibited, for example in pico-hydro-turbine for rural electrification. Since the smartphones are almost everywhere, they contribute to cost and environmental impact reductions in such solutions. The rest of the system is composed of reused components according to the principles of frugal innovation.

Interactive segmentation consists in building a pixel-wise partition of an image, into foreground and background regions, with the help of user inputs. Most state-of-the-art algorithms use scribble-based interactions to build foreground and background models, and very few of these work focus on the usability of the scribbling interaction. In this paper we study the outlining interaction, which is very intuitive to non-expert users on touch devices. We present an algorithm, built upon the existing GrabCut algorithm, which infers both foreground and background models out of a single outline. We conducted a user study on 20 participants to demonstrate the usability of this interaction, and its performance for the task of interactive segmentation.

This paper presents a multilevel analysis of 2D shapes and uses it to find similarities between the different parts of a shape. Such an analysis is important for many applications such as shape comparison, editing, and compression. Our robust and stable method decomposes a shape into parts, determines a parts hierarchy, and measures similarity between parts based on a salience measure on the medial axis, the Weighted Extended Distance Function, providing a multi-resolution partition of the shape that is stable across scale and articulation. Comparison with an extensive user study on the MPEG-7 database demonstrates that our geometric results are consistent with user perception.

A 3D bookmark in a networked virtual environment (NVE) provides a navigation aid, allowing the user to move quickly from its current viewpoint to a bookmarked viewpoint by simply clicking on the bookmark. In this paper, we first validate the positive impact that 3D bookmarks have in easing navigation in a 3D scene. Then, we show that, in the context of a NVE that streams content on demand from server to client, navigating with bookmarks leads to lower rendering quality at the bookmarked viewpoint, due to lower locality of data. We then investigate into how prefetching the 3D data at the bookmarks and precomputation of visible faces at the bookmarks help to improve the rendering quality.

In this paper we present the 2D Shape Structure database, a public, user-generated dataset of 2D shape decompositions into a hierarchy of shape parts with geometric relationships retained. It is the outcome of a large-scale user study obtained by crowdsourcing, involving over 1200 shapes in 70 shape classes, and 2861 participants. A total of 41,953 annotations has been collected with at least 24 annotations per shape. For each shape, user decompositions into main shape, one or more levels of parts, and a level of details are available. This database reinforces a philosophy that understanding shape structure as a whole, rather than in the separated categories of parts decomposition, parts hierarchy, and analysis of relationships between parts, is crucial for full shape understanding. We provide initial statistical explorations of the data to determine representative (“mean”) shape annotations and to determine the number of modes in the annotations. The primary goal of the paper is to make this rich and complex database openly available (through the website http://2dshapesstructure.github.io/index.html), providing the shape community with a ground truth of human perception of holistic shape structure.

This paper explores processing techniques to deal with noisy data in crowdsourced object segmentation tasks. We use the data collected with Click'n'Cut, an online interactive segmentation tool, and we perform several experiments towards improving the segmentation results. First, we introduce different superpixel-based techniques to filter users' traces, and assess their impact on the segmentation result. Second, we present different criteria to detect and discard the traces from potential bad users, resulting in a remarkable increase in performance. Finally, we show a novel superpixel-based segmentation algorithm which does not require any prior filtering and is based on weighting each user's contribution according to his/her level of expertise.

Reducing seek latency by predicting what the users will access is important for user experience, particularly during video browsing, where users seek frequently to skim through a video. Much existing research strived to predict user access pattern more accurately to improve the prefetching hit rate. This paper proposed a different approach whereby the prefetch hit rate is improved by biasing the users to seek to prefetched content with higher probability, through changing the video player user interface. Through a user study, we demonstrated that our player interface can lead to up to 4$\times$ more seeks to bookmarked segments and reduce seek latency by 40\%, compared to a video player interface commonly used today. The user study also showed that the user experience and the understanding of the video content when browsing is not compromised by the changes in seek behavior.

Videos are becoming more and more a tool of communication. There are how-to videos, people are discussing actions of others based on their recorded performance, e.g., in soccer, or they simply record videos of great moments and show them to friends and family. In this paper we focus on very specific how-to videos and present a novel, web based annotation tool, that combines (i) zoom, (ii) drawing, and (iii) temporal social bookmarking in video streams. Moreover, we present a short study on the usefulness of the tool to communicate general concepts of a specific video game based on a captured game session.

We consider an emerging situation where multiple cameras are filming the same event simultaneously from a diverse set of angles. The captured videos provide us with the multiple view geometry and an understanding of the 3D structure of the scene. We further extend this understanding by introducing the concept of 3D interest map in this paper. As most users naturally film what they find interesting from their respective viewpoints, the 3D structure can be annotated with the level of interest, naturally crowdsourced from the users. A 3D interest map can be understood as an extension of saliency maps in the 3D space that captures the semantics of the scene. We evaluate the idea of 3D interest maps on two real datasets, taken from the environment or the cameras that are equipped enough to have an estimation of the poses of cameras and a reasonable synchronization between them. We study two aspects of the 3D interest maps in our evaluation. First, by projecting them into 2D, we compare them to state-of-the-art saliency maps. Second, to demonstrate the usefulness of the 3D interest maps, we apply them to a video mashup system that automatically produces an edited video from one of the datasets.

In this demonstration, we show an automated mobile video mashup system that takes a set of videos filming the same scene as input, and generate an output mashup video consisting of temporally coherent clips selected from these input videos. The key difference of our system over the existing state-of-the-art is that it can generate virtual close-up shots and three camera operations: zooming in, zooming out, and panning, automatically. To achieve this, the system first computes the motion maps of the input videos and then determines a set of rectangles that correspond to highly interesting regions (in terms of motion). The choice of which shot types to use is done heuristically, ensuring diversity and coherency in the content presented in the mashup.

Online galleries of 3D models typically provide two ways to preview a model before the model is downloaded and viewed by the user: (i) by showing a set of thumbnail images of the 3D model taken from representative views (or keyviews); (ii) by showing a video of the 3D model as viewed from a moving virtual camera along a path determined by the content provider. We propose a third approach called preview streaming for mesh-based 3D objects: by streaming and showing parts of the mesh surfaces visible along the virtual camera path. This article focuses on the preview streaming architecture and framework and presents our investigation into how such a system would best handle network congestion effectively. We present three basic methods: (a) stop-and-wait , where the camera pauses until sufficient data is buffered; (b) reduce-speed , where the camera slows down in accordance to reduce network bandwidth; and (c) reduce-quality , where the camera continues to move at the same speed but fewer vertices are sent and displayed, leading to lower mesh quality. We further propose two advanced methods: (d) keyview-aware , which trades off mesh quality and camera speed appropriately depending on how close the current view is to the keyviews, and (e) adaptive-zoom , which improves visual quality by moving the virtual camera away from the original path. A user study reveals that our keyview-aware method is preferred over the basic methods. Moreover, the adaptive-zoom scheme compares favorably to the keyview-aware method, showing that path adaptation is a viable approach to handling bandwidth variation.

Publishers of 3D models online typically provide two ways to preview a model before the model is downloaded and viewed by the user: (i) by showing a set of thumbnail images of the 3D model taken from representative views (or keyviews); (ii) by showing a video of the 3D model as viewed from a moving virtual camera along a path determined by the content provider. We propose a third approach called preview streaming for mesh-based 3D object: by streaming and showing parts of the mesh surfaces visible along the virtual camera path. This paper focuses on the preview streaming architecture and framework, and presents our investigation into how such a system would best handle network congestion effectively. We study three basic methods: (a) stop-and-wait, where the camera pauses until sufficient data is buffered; (b) reduce-speed, where the camera slows down in accordance to reduce network bandwidth; and (c) reduce-quality, where the camera continues to move at the same speed but fewer vertices are sent and displayed, leading to lower mesh quality. We further propose a keyview-aware method that trades off mesh quality and camera speed appropriately depending on how close the current view is to the keyviews. A user study reveals that our keyview-aware method is preferred over the basic methods.

This paper introduces a new paradigm for interacting with zoomable video. Our interaction technique reduces the number of zooms and pans required by providing recommended viewports to the users, and replaces multiple zoom and pan actions with a simple click on the recommended viewport. The usefulness of our technique is visible in the quality of the recommended viewport, which needs to match the user intention, track movement in the scene, and properly frame the scene in the video. To this end, we propose a hybrid method where content analysis is complimented by the implicit feedback of a community of users in order to recommend viewports. We first compute preliminary sets of recommended viewports by analyzing the content of the video. These viewports allow tracking of moving objects in the scene, and are framed without violating basic aesthetic rules. To improve the relevance of the recommended viewports, we collect viewing statistics as users view a video, and use the viewports they select to reinforce the importance of certain recommendations and penalize others. New recommendations that are not previously recognized by content analysis may also emerge. The resulting recommended viewports converge towards the regions in the video that are relevant to users. A user study involving 70 participants shows that an user interface incorporating with our paradigm leads to more number of zooms, into more informative regions with fewer interactions.

We present a new user interface designed to allow easy yet effective zooming and panning into high-definition videos for playback on low resolution displays. Our system first applies state-of-the-art video analysis algorithms to detect salient regions of interest and recommends them to users. These recommendations help users to quickly identify important regions in the video and zoom into the regions with a single mouse click. The salient regions may moves according to the movement of track objects, further reducing the need for users to manually pan to track an object of interests. To further improve the relevance of the recommended regions, users' interactions are logged and analyzed. The actual regions selected and viewed by users serve as a feedback and is integrated into the system to improve the recommendations. We have implemented a Web-based version of the user interface, running on modern browsers supporting HTML5. We describe the algorithms and optimizations used to implement and improve the system.

Screen size and display resolution limit the experience of watching videos on mobile devices. The viewing experience can be improved by determining important or interesting regions within the video (called regions of interest, or ROIs) and displaying only the ROIs to the viewer. Previous work focuses on analyzing the video content using visual attention model to infer the ROIs. Such content-based technique, however, has limitations. In this paper, we propose an alternative paradigm to infer ROIs from a video. We crowdsource from a large number of users through their implicit viewing behavior using a zoom and pan interface, and infer the ROIs from their collective wisdom. A retargeted video, consisting of relevant shots determined from historical users behavior, can be automatically generated and replayed to subsequent users who would prefer a less interactive viewing experience. This paper presents how we collect the user traces, infer the ROIs and their dynamics, group the ROIs into shots, and automatically reframe those shots to improve the aesthetics of the video. A user study with 48 participants shows that our automatically retargeted video is of comparable quality to one handcrafted by an expert user

Streaming of an arbitrary region of interest (RoI) from a high resolution video is essential to supporting zooming and panning within a video stream. This paper explores two methods for RoI-based streaming, referring to them as tiled streaming and monolithic streaming. Tiled streaming partitions video frames into grid of tiles and encodes each tile as an independently decodable stream. Monolithic streaming applies to video encoded using off-the-shelf encoder, and relies on a pre-computed dependency information to send the necessary bits for the RoI. We evaluated these two methods in terms of bandwidth efficiency, storage requirement, and computational costs under different video encoding parameters. Experimental results show that bandwidth efficiency of tiled streams for RoI-based streaming reduces when tile size increases, despite improvement in compression efficiency. In the case of monolithic streams, use of a larger motion vector range coupled with careful run-time optimization can still improve the bandwidth efficiency, despite an increase in motion vector dependency.