Esa Rahtu

Dingding Cai, Janne Heikkilä, Esa Rahtu

This paper presents an efficient symmetry-agnostic and correspondence-free framework, referred to as SC6D, for 6D object pose estimation from a single monocular RGB image. SC6D requires neither the 3D CAD model of the object nor any prior knowledge of the symmetries. The pose estimation is decomposed into three sub-tasks: a) object 3D rotation representation learning and matching; b) estimation of the 2D location of the object center; and c) scale-invariant distance estimation (the translation along the z-axis) via classification. SC6D is evaluated on three benchmark datasets, T-LESS, YCB-V, and ITODD, and results in state-of-the-art performance on the T-LESS dataset. Moreover, SC6D is computationally much more efficient than the previous state-of-the-art method SurfEmb. The implementation and pre-trained models are publicly available at https://github.com/dingdingcai/SC6D-pose.

Vladimir Iashin, Weidi Xie, Esa Rahtu et al.

The objective of this paper is audio-visual synchronisation of general videos 'in the wild'. For such videos, the events that may be harnessed for synchronisation cues may be spatially small and may occur only infrequently during a many seconds-long video clip, i.e. the synchronisation signal is 'sparse in space and time'. This contrasts with the case of synchronising videos of talking heads, where audio-visual correspondence is dense in both time and space. We make four contributions: (i) in order to handle longer temporal sequences required for sparse synchronisation signals, we design a multi-modal transformer model that employs 'selectors' to distil the long audio and visual streams into small sequences that are then used to predict the temporal offset between streams. (ii) We identify artefacts that can arise from the compression codecs used for audio and video and can be used by audio-visual models in training to artificially solve the synchronisation task. (iii) We curate a dataset with only sparse in time and space synchronisation signals; and (iv) the effectiveness of the proposed model is shown on both dense and sparse datasets quantitatively and qualitatively. Project page: v-iashin.github.io/SparseSync

Mayu Otani, Riku Togashi, Yu Sawai et al.

Human evaluation is critical for validating the performance of text-to-image generative models, as this highly cognitive process requires deep comprehension of text and images. However, our survey of 37 recent papers reveals that many works rely solely on automatic measures (e.g., FID) or perform poorly described human evaluations that are not reliable or repeatable. This paper proposes a standardized and well-defined human evaluation protocol to facilitate verifiable and reproducible human evaluation in future works. In our pilot data collection, we experimentally show that the current automatic measures are incompatible with human perception in evaluating the performance of the text-to-image generation results. Furthermore, we provide insights for designing human evaluation experiments reliably and conclusively. Finally, we make several resources publicly available to the community to facilitate easy and fast implementations.

Dingding Cai, Janne Heikkilä, Esa Rahtu

This paper proposes a universal framework, called OVE6D, for model-based 6D object pose estimation from a single depth image and a target object mask. Our model is trained using purely synthetic data rendered from ShapeNet, and, unlike most of the existing methods, it generalizes well on new real-world objects without any fine-tuning. We achieve this by decomposing the 6D pose into viewpoint, in-plane rotation around the camera optical axis and translation, and introducing novel lightweight modules for estimating each component in a cascaded manner. The resulting network contains less than 4M parameters while demonstrating excellent performance on the challenging T-LESS and Occluded LINEMOD datasets without any dataset-specific training. We show that OVE6D outperforms some contemporary deep learning-based pose estimation methods specifically trained for individual objects or datasets with real-world training data. The implementation and the pre-trained model will be made publicly available.

Juan Pablo Lagos, Esa Rahtu

Holistic scene understanding is pivotal for the performance of autonomous machines. In this paper we propose a new end-to-end model for performing semantic segmentation and depth completion jointly. The vast majority of recent approaches have developed semantic segmentation and depth completion as independent tasks. Our approach relies on RGB and sparse depth as inputs to our model and produces a dense depth map and the corresponding semantic segmentation image. It consists of a feature extractor, a depth completion branch, a semantic segmentation branch and a joint branch which further processes semantic and depth information altogether. The experiments done on Virtual KITTI 2 dataset, demonstrate and provide further evidence, that combining both tasks, semantic segmentation and depth completion, in a multi-task network can effectively improve the performance of each task. Code is available at https://github.com/juanb09111/semantic depth.

Mayu Otani, Riku Togashi, Yuta Nakashima et al.

Mean Average Precision (mAP) is the primary evaluation measure for object detection. Although object detection has a broad range of applications, mAP evaluates detectors in terms of the performance of ranked instance retrieval. Such the assumption for the evaluation task does not suit some downstream tasks. To alleviate the gap between downstream tasks and the evaluation scenario, we propose Optimal Correction Cost (OC-cost), which assesses detection accuracy at image level. OC-cost computes the cost of correcting detections to ground truths as a measure of accuracy. The cost is obtained by solving an optimal transportation problem between the detections and the ground truths. Unlike mAP, OC-cost is designed to penalize false positive and false negative detections properly, and every image in a dataset is treated equally. Our experimental result validates that OC-cost has better agreement with human preference than a ranking-based measure, i.e., mAP for a single image. We also show that detectors' rankings by OC-cost are more consistent on different data splits than mAP. Our goal is not to replace mAP with OC-cost but provide an additional tool to evaluate detectors from another aspect. To help future researchers and developers choose a target measure, we provide a series of experiments to clarify how mAP and OC-cost differ.

Xuqian Ren, Wenjia Wang, Dingding Cai et al.

Metaverse technologies demand accurate, real-time, and immersive modeling on consumer-grade hardware for both non-human perception (e.g., drone/robot/autonomous car navigation) and immersive technologies like AR/VR, requiring both structural accuracy and photorealism. However, there exists a knowledge gap in how to apply geometric reconstruction and photorealism modeling (novel view synthesis) in a unified framework. To address this gap and promote the development of robust and immersive modeling and rendering with consumer-grade devices, we propose a real-world Multi-Sensor Hybrid Room Dataset (MuSHRoom). Our dataset presents exciting challenges and requires state-of-the-art methods to be cost-effective, robust to noisy data and devices, and can jointly learn 3D reconstruction and novel view synthesis instead of treating them as separate tasks, making them ideal for real-world applications. We benchmark several famous pipelines on our dataset for joint 3D mesh reconstruction and novel view synthesis. Our dataset and benchmark show great potential in promoting the improvements for fusing 3D reconstruction and high-quality rendering in a robust and computationally efficient end-to-end fashion. The dataset and code are available at the project website: https://xuqianren.github.io/publications/MuSHRoom/.

Juan Lagos, Esa Rahtu

Understanding 3D environments semantically is pivotal in autonomous driving applications where multiple computer vision tasks are involved. Multi-task models provide different types of outputs for a given scene, yielding a more holistic representation while keeping the computational cost low. We propose a multi-task model for panoptic segmentation and depth completion using RGB images and sparse depth maps. Our model successfully predicts fully dense depth maps and performs semantic segmentation, instance segmentation, and panoptic segmentation for every input frame. Extensive experiments were done on the Virtual KITTI 2 dataset and we demonstrate that our model solves multiple tasks, without a significant increase in computational cost, while keeping high accuracy performance. Code is available at https://github.com/juanb09111/PanDepth.git

Lassi Raatikainen, Esa Rahtu

The objective of this paper is to assess the quality of explanation heatmaps for image classification tasks. To assess the quality of explainability methods, we approach the task through the lens of accuracy and stability. In this work, we make the following contributions. Firstly, we introduce the Weighting Game, which measures how much of a class-guided explanation is contained within the correct class' segmentation mask. Secondly, we introduce a metric for explanation stability, using zooming/panning transformations to measure differences between saliency maps with similar contents. Quantitative experiments are produced, using these new metrics, to evaluate the quality of explanations provided by commonly used CAM methods. The quality of explanations is also contrasted between different model architectures, with findings highlighting the need to consider model architecture when choosing an explainability method.

Phong Nguyen-Ha, Lam Huynh, Esa Rahtu et al.

We present CG-NeRF, a cascade and generalizable neural radiance fields method for view synthesis. Recent generalizing view synthesis methods can render high-quality novel views using a set of nearby input views. However, the rendering speed is still slow due to the nature of uniformly-point sampling of neural radiance fields. Existing scene-specific methods can train and render novel views efficiently but can not generalize to unseen data. Our approach addresses the problems of fast and generalizing view synthesis by proposing two novel modules: a coarse radiance fields predictor and a convolutional-based neural renderer. This architecture infers consistent scene geometry based on the implicit neural fields and renders new views efficiently using a single GPU. We first train CG-NeRF on multiple 3D scenes of the DTU dataset, and the network can produce high-quality and accurate novel views on unseen real and synthetic data using only photometric losses. Moreover, our method can leverage a denser set of reference images of a single scene to produce accurate novel views without relying on additional explicit representations and still maintains the high-speed rendering of the pre-trained model. Experimental results show that CG-NeRF outperforms state-of-the-art generalizable neural rendering methods on various synthetic and real datasets.

Dingding Cai, Janne Heikkilä, Esa Rahtu

Acquiring labeled 6D poses from real images is an expensive and time-consuming task. Though massive amounts of synthetic RGB images are easy to obtain, the models trained on them suffer from noticeable performance degradation due to the synthetic-to-real domain gap. To mitigate this degradation, we propose a practical self-supervised domain adaptation approach that takes advantage of real RGB(-D) data without needing real pose labels. We first pre-train the model with synthetic RGB images and then utilize real RGB(-D) images to fine-tune the pre-trained model. The fine-tuning process is self-supervised by the RGB-based pose-aware consistency and the depth-guided object distance pseudo-label, which does not require the time-consuming online differentiable rendering. We build our domain adaptation method based on the recent pose estimator SC6D and evaluate it on the YCB-Video dataset. We experimentally demonstrate that our method achieves comparable performance against its fully-supervised counterpart while outperforming existing state-of-the-art approaches.

Janne Mustaniemi, Juho Kannala, Esa Rahtu et al.

Various datasets have been proposed for simultaneous localization and mapping (SLAM) and related problems. Existing datasets often include small environments, have incomplete ground truth, or lack important sensor data, such as depth and infrared images. We propose an easy-to-use framework for acquiring building-scale 3D reconstruction using a consumer depth camera. Unlike complex and expensive acquisition setups, our system enables crowd-sourcing, which can greatly benefit data-hungry algorithms. Compared to similar systems, we utilize raw depth maps for odometry computation and loop closure refinement which results in better reconstructions. We acquire a building-scale 3D dataset (BS3D) and demonstrate its value by training an improved monocular depth estimation model. As a unique experiment, we benchmark visual-inertial odometry methods using both color and active infrared images.

Lam Huynh, Esa Rahtu, Jiri Matas et al.

We present LDP, a lightweight dense prediction neural architecture search (NAS) framework. Starting from a pre-defined generic backbone, LDP applies the novel Assisted Tabu Search for efficient architecture exploration. LDP is fast and suitable for various dense estimation problems, unlike previous NAS methods that are either computational demanding or deployed only for a single subtask. The performance of LPD is evaluated on monocular depth estimation, semantic segmentation, and image super-resolution tasks on diverse datasets, including NYU-Depth-v2, KITTI, Cityscapes, COCO-stuff, DIV2K, Set5, Set14, BSD100, Urban100. Experiments show that the proposed framework yields consistent improvements on all tested dense prediction tasks, while being $5\%-315\%$ more compact in terms of the number of model parameters than prior arts.

Leevi Raivio, Esa Rahtu

Real-time holistic scene understanding would allow machines to interpret their surrounding in a much more detailed manner than is currently possible. While panoptic image segmentation methods have brought image segmentation closer to this goal, this information has to be described relative to the 3D environment for the machine to be able to utilise it effectively. In this paper, we investigate methods for sequentially reconstructing static environments from panoptic image segmentations in 3D. We specifically target real-time operation: the algorithm must process data strictly online and be able to run at relatively fast frame rates. Additionally, the method should be scalable for environments large enough for practical applications. By applying a simple but powerful data-association algorithm, we outperform earlier similar works when operating purely online. Our method is also capable of reaching frame-rates high enough for real-time applications and is scalable to larger environments as well. Source code and further demonstrations are released to the public at: \url{https://tutvision.github.io/Online-Panoptic-3D/}

Ilpo Viertola, Vladimir Iashin, Esa Rahtu

We introduce V-AURA, the first autoregressive model to achieve high temporal alignment and relevance in video-to-audio generation. V-AURA uses a high-framerate visual feature extractor and a cross-modal audio-visual feature fusion strategy to capture fine-grained visual motion events and ensure precise temporal alignment. Additionally, we propose VisualSound, a benchmark dataset with high audio-visual relevance. VisualSound is based on VGGSound, a video dataset consisting of in-the-wild samples extracted from YouTube. During the curation, we remove samples where auditory events are not aligned with the visual ones. V-AURA outperforms current state-of-the-art models in temporal alignment and semantic relevance while maintaining comparable audio quality. Code, samples, VisualSound and models are available at https://v-aura.notion.site

Lingyu Zhu, Esa Rahtu, Hang Zhao

This paper focuses on perceiving and navigating 3D environments using echoes and RGB image. In particular, we perform depth estimation by fusing RGB image with echoes, received from multiple orientations. Unlike previous works, we go beyond the field of view of the RGB and estimate dense depth maps for substantially larger parts of the environment. We show that the echoes provide holistic and in-expensive information about the 3D structures complementing the RGB image. Moreover, we study how echoes and the wide field-of-view depth maps can be utilised in robot navigation. We compare the proposed methods against recent baselines using two sets of challenging realistic 3D environments: Replica and Matterport3D. The implementation and pre-trained models will be made publicly available.

Mostafa Mansour, Ahmed Abdelsalam, Ari Happonen et al.

Recent advancements in monocular neural depth estimation, particularly those achieved by the UniDepth network, have prompted the investigation of integrating UniDepth within a Gaussian splatting framework for monocular SLAM. This study presents UDGS-SLAM, a novel approach that eliminates the necessity of RGB-D sensors for depth estimation within Gaussian splatting framework. UDGS-SLAM employs statistical filtering to ensure local consistency of the estimated depth and jointly optimizes camera trajectory and Gaussian scene representation parameters. The proposed method achieves high-fidelity rendered images and low ATERMSE of the camera trajectory. The performance of UDGS-SLAM is rigorously evaluated using the TUM RGB-D dataset and benchmarked against several baseline methods, demonstrating superior performance across various scenarios. Additionally, an ablation study is conducted to validate design choices and investigate the impact of different network backbone encoders on system performance.

Juan Lagos, Urho Lempiö, Esa Rahtu

While the availability of large and diverse datasets has contributed to significant breakthroughs in autonomous driving and indoor applications, forestry applications are still lagging behind and new forest datasets would most certainly contribute to achieving significant progress in the development of data-driven methods for forest-like scenarios. This paper introduces a forest dataset called \textit{FinnWoodlands}, which consists of RGB stereo images, point clouds, and sparse depth maps, as well as ground truth manual annotations for semantic, instance, and panoptic segmentation. \textit{FinnWoodlands} comprises a total of 4226 objects manually annotated, out of which 2562 objects (60.6\%) correspond to tree trunks classified into three different instance categories, namely "Spruce Tree", "Birch Tree", and "Pine Tree". Besides tree trunks, we also annotated "Obstacles" objects as instances as well as the semantic stuff classes "Lake", "Ground", and "Track". Our dataset can be used in forestry applications where a holistic representation of the environment is relevant. We provide an initial benchmark using three models for instance segmentation, panoptic segmentation, and depth completion, and illustrate the challenges that such unstructured scenarios introduce.

Xuqian Ren, Matias Turkulainen, Jiepeng Wang et al.

Geometric priors are often used to enhance 3D reconstruction. With many smartphones featuring low-resolution depth sensors and the prevalence of off-the-shelf monocular geometry estimators, incorporating geometric priors as regularization signals has become common in 3D vision tasks. However, the accuracy of depth estimates from mobile devices is typically poor for highly detailed geometry, and monocular estimators often suffer from poor multi-view consistency and precision. In this work, we propose an approach for joint surface depth and normal refinement of Gaussian Splatting methods for accurate 3D reconstruction of indoor scenes. We develop supervision strategies that adaptively filters low-quality depth and normal estimates by comparing the consistency of the priors during optimization. We mitigate regularization in regions where prior estimates have high uncertainty or ambiguities. Our filtering strategy and optimization design demonstrate significant improvements in both mesh estimation and novel-view synthesis for both 3D and 2D Gaussian Splatting-based methods on challenging indoor room datasets. Furthermore, we explore the use of alternative meshing strategies for finer geometry extraction. We develop a scale-aware meshing strategy inspired by TSDF and octree-based isosurface extraction, which recovers finer details from Gaussian models compared to other commonly used open-source meshing tools. Our code is released in https://xuqianren.github.io/ags_mesh_website/.

Farid Alijani, Esa Rahtu

In this paper, we present a comprehensive study on the utility of deep convolutional neural networks with two state-of-the-art pooling layers which are placed after convolutional layers and fine-tuned in an end-to-end manner for visual place recognition task in challenging conditions, including seasonal and illumination variations. We compared extensively the performance of deep learned global features with three different loss functions, e.g. triplet, contrastive and ArcFace, for learning the parameters of the architectures in terms of fraction of the correct matches during deployment. To verify effectiveness of our results, we utilized two real world datasets in place recognition, both indoor and outdoor. Our investigation demonstrates that fine tuning architectures with ArcFace loss in an end-to-end manner outperforms other two losses by approximately 1~4% in outdoor and 1~2% in indoor datasets, given certain thresholds, for the visual place recognition tasks.

Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov et al.

High-fidelity 3D reconstruction of common indoor scenes is crucial for VR and AR applications. 3D Gaussian splatting, a novel differentiable rendering technique, has achieved state-of-the-art novel view synthesis results with high rendering speeds and relatively low training times. However, its performance on scenes commonly seen in indoor datasets is poor due to the lack of geometric constraints during optimization. In this work, we explore the use of readily accessible geometric cues to enhance Gaussian splatting optimization in challenging, ill-posed, and textureless scenes. We extend 3D Gaussian splatting with depth and normal cues to tackle challenging indoor datasets and showcase techniques for efficient mesh extraction. Specifically, we regularize the optimization procedure with depth information, enforce local smoothness of nearby Gaussians, and use off-the-shelf monocular networks to achieve better alignment with the true scene geometry. We propose an adaptive depth loss based on the gradient of color images, improving depth estimation and novel view synthesis results over various baselines. Our simple yet effective regularization technique enables direct mesh extraction from the Gaussian representation, yielding more physically accurate reconstructions of indoor scenes.

Bishwo Adhikari, Xingyang Ni, Esa Rahtu et al.

Facial analysis is an active research area in computer vision, with many practical applications. Most of the existing studies focus on addressing one specific task and maximizing its performance. For a complete facial analysis system, one needs to solve these tasks efficiently to ensure a smooth experience. In this work, we present a system-level design of a real-time facial analysis system. With a collection of deep neural networks for object detection, classification, and regression, the system recognizes age, gender, facial expression, and facial similarity for each person that appears in the camera view. We investigate the parallelization and interplay of individual tasks. Results on common off-the-shelf architecture show that the system's accuracy is comparable to the state-of-the-art methods, and the recognition speed satisfies real-time requirements. Moreover, we propose a multitask network for jointly predicting the first three attributes, i.e., age, gender, and facial expression. Source code and trained models are available at https://github.com/mahehu/TUT-live-age-estimator.

Xingyang Ni, Heikki Huttunen, Esa Rahtu

In this study, we analyze model inversion attacks with only two assumptions: feature vectors of user data are known, and a black-box API for inference is provided. On the one hand, limitations of existing studies are addressed by opting for a more practical setting. Experiments have been conducted on state-of-the-art models in person re-identification, and two attack scenarios (i.e., recognizing auxiliary attributes and reconstructing user data) are investigated. Results show that an adversary could successfully infer sensitive information even under severe constraints. On the other hand, it is advisable to encrypt feature vectors, especially for a machine learning model in production. As an alternative to traditional encryption methods such as AES, a simple yet effective method termed ShuffleBits is presented. More specifically, the binary sequence of each floating-point number gets shuffled. Deployed using the one-time pad scheme, it serves as a plug-and-play module that is applicable to any neural network, and the resulting model directly outputs deep features in encrypted form. Source code is publicly available at https://github.com/nixingyang/ShuffleBits.

Otto Seiskari, Pekka Rantalankila, Juho Kannala et al.

We present HybVIO, a novel hybrid approach for combining filtering-based visual-inertial odometry (VIO) with optimization-based SLAM. The core of our method is highly robust, independent VIO with improved IMU bias modeling, outlier rejection, stationarity detection, and feature track selection, which is adjustable to run on embedded hardware. Long-term consistency is achieved with a loosely-coupled SLAM module. In academic benchmarks, our solution yields excellent performance in all categories, especially in the real-time use case, where we outperform the current state-of-the-art. We also demonstrate the feasibility of VIO for vehicular tracking on consumer-grade hardware using a custom dataset, and show good performance in comparison to current commercial VISLAM alternatives. An open-source implementation of the HybVIO method is available at https://github.com/SpectacularAI/HybVIO

Xingyang Ni, Esa Rahtu

Since neural networks are data-hungry, incorporating data augmentation in training is a widely adopted technique that enlarges datasets and improves generalization. On the other hand, aggregating predictions of multiple augmented samples (i.e., test-time augmentation) could boost performance even further. In the context of person re-identification models, it is common practice to extract embeddings for both the original images and their horizontally flipped variants. The final representation is the mean of the aforementioned feature vectors. However, such scheme results in a gap between training and inference, i.e., the mean feature vectors calculated in inference are not part of the training pipeline. In this study, we devise the FlipReID structure with the flipping loss to address this issue. More specifically, models using the FlipReID structure are trained on the original images and the flipped images simultaneously, and incorporating the flipping loss minimizes the mean squared error between feature vectors of corresponding image pairs. Extensive experiments show that our method brings consistent improvements. In particular, we set a new record for MSMT17 which is the largest person re-identification dataset. The source code is available at https://github.com/nixingyang/FlipReID.

Lingyu Zhu, Esa Rahtu

Visual sound source separation aims at identifying sound components from a given sound mixture with the presence of visual cues. Prior works have demonstrated impressive results, but with the expense of large multi-stage architectures and complex data representations (e.g. optical flow trajectories). In contrast, we study simple yet efficient models for visual sound separation using only a single video frame. Furthermore, our models are able to exploit the information of the sound source category in the separation process. To this end, we propose two models where we assume that i) the category labels are available at the training time, or ii) we know if the training sample pairs are from the same or different category. The experiments with the MUSIC dataset show that our model obtains comparable or better performance compared to several recent baseline methods. The code is available at https://github.com/ly-zhu/Leveraging-Category-Information-for-Single-Frame-Visual-Sound-Source-Separation

Vladimir Iashin, Esa Rahtu

Dense video captioning aims to localize and describe important events in untrimmed videos. Existing methods mainly tackle this task by exploiting only visual features, while completely neglecting the audio track. Only a few prior works have utilized both modalities, yet they show poor results or demonstrate the importance on a dataset with a specific domain. In this paper, we introduce Bi-modal Transformer which generalizes the Transformer architecture for a bi-modal input. We show the effectiveness of the proposed model with audio and visual modalities on the dense video captioning task, yet the module is capable of digesting any two modalities in a sequence-to-sequence task. We also show that the pre-trained bi-modal encoder as a part of the bi-modal transformer can be used as a feature extractor for a simple proposal generation module. The performance is demonstrated on a challenging ActivityNet Captions dataset where our model achieves outstanding performance. The code is available: v-iashin.github.io/bmt

Vladimir Iashin, Esa Rahtu

Dense video captioning is a task of localizing interesting events from an untrimmed video and producing textual description (captions) for each localized event. Most of the previous works in dense video captioning are solely based on visual information and completely ignore the audio track. However, audio, and speech, in particular, are vital cues for a human observer in understanding an environment. In this paper, we present a new dense video captioning approach that is able to utilize any number of modalities for event description. Specifically, we show how audio and speech modalities may improve a dense video captioning model. We apply automatic speech recognition (ASR) system to obtain a temporally aligned textual description of the speech (similar to subtitles) and treat it as a separate input alongside video frames and the corresponding audio track. We formulate the captioning task as a machine translation problem and utilize recently proposed Transformer architecture to convert multi-modal input data into textual descriptions. We demonstrate the performance of our model on ActivityNet Captions dataset. The ablation studies indicate a considerable contribution from audio and speech components suggesting that these modalities contain substantial complementary information to video frames. Furthermore, we provide an in-depth analysis of the ActivityNet Caption results by leveraging the category tags obtained from original YouTube videos. Code is publicly available: github.com/v-iashin/MDVC

Hamed R. Tavakoli, Ali Borji, Esa Rahtu et al.

This paper studies audio-visual deep saliency prediction. It introduces a conceptually simple and effective Deep Audio-Visual Embedding for dynamic saliency prediction dubbed ``DAVE" in conjunction with our efforts towards building an Audio-Visual Eye-tracking corpus named ``AVE". Despite existing a strong relation between auditory and visual cues for guiding gaze during perception, video saliency models only consider visual cues and neglect the auditory information that is ubiquitous in dynamic scenes. Here, we investigate the applicability of audio cues in conjunction with visual ones in predicting saliency maps using deep neural networks. To this end, the proposed model is intentionally designed to be simple. Two baseline models are developed on the same architecture which consists of an encoder-decoder. The encoder projects the input into a feature space followed by a decoder that infers saliency. We conduct an extensive analysis on different modalities and various aspects of multi-model dynamic saliency prediction. Our results suggest that (1) audio is a strong contributing cue for saliency prediction, (2) salient visible sound-source is the natural cause of the superiority of our Audio-Visual model, (3) richer feature representations for the input space leads to more powerful predictions even in absence of more sophisticated saliency decoders, and (4) Audio-Visual model improves over 53.54\% of the frames predicted by the best Visual model (our baseline). Our endeavour demonstrates that audio is an important cue that boosts dynamic video saliency prediction and helps models to approach human performance. The code is available at https://github.com/hrtavakoli/DAVE

Vladimir Iashin, Weidi Xie, Esa Rahtu et al.

Our objective is audio-visual synchronization with a focus on 'in-the-wild' videos, such as those on YouTube, where synchronization cues can be sparse. Our contributions include a novel audio-visual synchronization model, and training that decouples feature extraction from synchronization modelling through multi-modal segment-level contrastive pre-training. This approach achieves state-of-the-art performance in both dense and sparse settings. We also extend synchronization model training to AudioSet a million-scale 'in-the-wild' dataset, investigate evidence attribution techniques for interpretability, and explore a new capability for synchronization models: audio-visual synchronizability.

Otto Seiskari, Jerry Ylilammi, Valtteri Kaatrasalo et al.

High-quality scene reconstruction and novel view synthesis based on Gaussian Splatting (3DGS) typically require steady, high-quality photographs, often impractical to capture with handheld cameras. We present a method that adapts to camera motion and allows high-quality scene reconstruction with handheld video data suffering from motion blur and rolling shutter distortion. Our approach is based on detailed modelling of the physical image formation process and utilizes velocities estimated using visual-inertial odometry (VIO). Camera poses are considered non-static during the exposure time of a single image frame and camera poses are further optimized in the reconstruction process. We formulate a differentiable rendering pipeline that leverages screen space approximation to efficiently incorporate rolling-shutter and motion blur effects into the 3DGS framework. Our results with both synthetic and real data demonstrate superior performance in mitigating camera motion over existing methods, thereby advancing 3DGS in naturalistic settings.

Dingding Cai, Janne Heikkilä, Esa Rahtu

This paper introduces GS-Pose, a unified framework for localizing and estimating the 6D pose of novel objects. GS-Pose begins with a set of posed RGB images of a previously unseen object and builds three distinct representations stored in a database. At inference, GS-Pose operates sequentially by locating the object in the input image, estimating its initial 6D pose using a retrieval approach, and refining the pose with a render-and-compare method. The key insight is the application of the appropriate object representation at each stage of the process. In particular, for the refinement step, we leverage 3D Gaussian splatting, a novel differentiable rendering technique that offers high rendering speed and relatively low optimization time. Off-the-shelf toolchains and commodity hardware, such as mobile phones, can be used to capture new objects to be added to the database. Extensive evaluations on the LINEMOD and OnePose-LowTexture datasets demonstrate excellent performance, establishing the new state-of-the-art. Project page: https://dingdingcai.github.io/gs-pose.

Ulas Gunes, Matias Turkulainen, Xuqian Ren et al.

The development of large-scale 3D scene reconstruction and novel view synthesis methods mostly rely on datasets comprising perspective images with narrow fields of view (FoV). While effective for small-scale scenes, these datasets require large image sets and extensive structure-from-motion (SfM) processing, limiting scalability. To address this, we introduce a fisheye image dataset tailored for scene reconstruction tasks. Using dual 200-degree fisheye lenses, our dataset provides full 360-degree coverage of 5 indoor and 5 outdoor scenes. Each scene has sparse SfM point clouds and precise LIDAR-derived dense point clouds that can be used as geometric ground-truth, enabling robust benchmarking under challenging conditions such as occlusions and reflections. While the baseline experiments focus on vanilla Gaussian Splatting and NeRF based Nerfacto methods, the dataset supports diverse approaches for scene reconstruction, novel view synthesis, and image-based rendering.

Ilpo Viertola, Vladimir Iashin, Esa Rahtu

Existing multimodal audio generation models often lack precise user control, which limits their applicability in professional Foley workflows. In particular, these models focus on the entire video and do not provide precise methods for prioritizing a specific object within a scene, generating unnecessary background sounds, or focusing on the wrong objects. To address this gap, we introduce the novel task of video object segmentation-aware audio generation, which explicitly conditions sound synthesis on object-level segmentation maps. We present SAGANet, a new multimodal generative model that enables controllable audio generation by leveraging visual segmentation masks along with video and textual cues. Our model provides users with fine-grained and visually localized control over audio generation. To support this task and further research on segmentation-aware Foley, we propose Segmented Music Solos, a benchmark dataset of musical instrument performance videos with segmentation information. Our method demonstrates substantial improvements over current state-of-the-art methods and sets a new standard for controllable, high-fidelity Foley synthesis. Code, samples, and Segmented Music Solos are available at https://saganet.notion.site

Ulas Gunes, Matias Turkulainen, Juho Kannala et al.

We present the first evaluation of fisheye-based 3D Gaussian Splatting methods, Fisheye-GS and 3DGUT, on real images with fields of view exceeding 180 degree. Our study covers both indoor and outdoor scenes captured with 200 degree fisheye cameras and analyzes how each method handles extreme distortion in real world settings. We evaluate performance under varying fields of view (200 degree, 160 degree, and 120 degree) to study the tradeoff between peripheral distortion and spatial coverage. Fisheye-GS benefits from field of view (FoV) reduction, particularly at 160 degree, while 3DGUT remains stable across all settings and maintains high perceptual quality at the full 200 degree view. To address the limitations of SfM-based initialization, which often fails under strong distortion, we also propose a depth-based strategy using UniK3D predictions from only 2-3 fisheye images per scene. Although UniK3D is not trained on real fisheye data, it produces dense point clouds that enable reconstruction quality on par with SfM, even in difficult scenes with fog, glare, or sky. Our results highlight the practical viability of fisheye-based 3DGS methods for wide-angle 3D reconstruction from sparse and distortion-heavy image inputs.

Jukka I. Ahonen, Nam Le, Honglei Zhang et al.

The recent progress in artificial intelligence has led to an ever-increasing usage of images and videos by machine analysis algorithms, mainly neural networks. Nonetheless, compression, storage and transmission of media have traditionally been designed considering human beings as the viewers of the content. Recent research on image and video coding for machine analysis has progressed mainly in two almost orthogonal directions. The first is represented by end-to-end (E2E) learned codecs which, while offering high performance on image coding, are not yet on par with state-of-the-art conventional video codecs and lack interoperability. The second direction considers using the Versatile Video Coding (VVC) standard or any other conventional video codec (CVC) together with pre- and post-processing operations targeting machine analysis. While the CVC-based methods benefit from interoperability and broad hardware and software support, the machine task performance is often lower than the desired level, particularly in low bitrates. This paper proposes a hybrid codec for machines called NN-VVC, which combines the advantages of an E2E-learned image codec and a CVC to achieve high performance in both image and video coding for machines. Our experiments show that the proposed system achieved up to -43.20% and -26.8% Bjøntegaard Delta rate reduction over VVC for image and video data, respectively, when evaluated on multiple different datasets and machine vision tasks. To the best of our knowledge, this is the first research paper showing a hybrid video codec that outperforms VVC on multiple datasets and multiple machine vision tasks.

Nam Le, Honglei Zhang, Francesco Cricri et al.

Image coding for machines (ICM) aims at reducing the bitrate required to represent an image while minimizing the drop in machine vision analysis accuracy. In many use cases, such as surveillance, it is also important that the visual quality is not drastically deteriorated by the compression process. Recent works on using neural network (NN) based ICM codecs have shown significant coding gains against traditional methods; however, the decompressed images, especially at low bitrates, often contain checkerboard artifacts. We propose an effective decoder finetuning scheme based on adversarial training to significantly enhance the visual quality of ICM codecs, while preserving the machine analysis accuracy, without adding extra bitcost or parameters at the inference phase. The results show complete removal of the checkerboard artifacts at the negligible cost of -1.6% relative change in task performance score. In the cases where some amount of artifacts is tolerable, such as when machine consumption is the primary target, this technique can enhance both pixel-fidelity and feature-fidelity scores without losing task performance.

Riku Togashi, Mayu Otani, Yuta Nakashima et al.

Evaluation measures have a crucial impact on the direction of research. Therefore, it is of utmost importance to develop appropriate and reliable evaluation measures for new applications where conventional measures are not well suited. Video Moment Retrieval (VMR) is one such application, and the current practice is to use R@$K,θ$ for evaluating VMR systems. However, this measure has two disadvantages. First, it is rank-insensitive: It ignores the rank positions of successfully localised moments in the top-$K$ ranked list by treating the list as a set. Second, it binarizes the Intersection over Union (IoU) of each retrieved video moment using the threshold $θ$ and thereby ignoring fine-grained localisation quality of ranked moments. We propose an alternative measure for evaluating VMR, called Average Max IoU (AxIoU), which is free from the above two problems. We show that AxIoU satisfies two important axioms for VMR evaluation, namely, \textbf{Invariance against Redundant Moments} and \textbf{Monotonicity with respect to the Best Moment}, and also that R@$K,θ$ satisfies the first axiom only. We also empirically examine how AxIoU agrees with R@$K,θ$, as well as its stability with respect to change in the test data and human-annotated temporal boundaries.

Nannan Zou, Honglei Zhang, Francesco Cricri et al.

Neural image coding represents now the state-of-the-art image compression approach. However, a lot of work is still to be done in the video domain. In this work, we propose an end-to-end learned video codec that introduces several architectural novelties as well as training novelties, revolving around the concepts of adaptation and attention. Our codec is organized as an intra-frame codec paired with an inter-frame codec. As one architectural novelty, we propose to train the inter-frame codec model to adapt the motion estimation process based on the resolution of the input video. A second architectural novelty is a new neural block that combines concepts from split-attention based neural networks and from DenseNets. Finally, we propose to overfit a set of decoder-side multiplicative parameters at inference time. Through ablation studies and comparisons to prior art, we show the benefits of our proposed techniques in terms of coding gains. We compare our codec to VVC/H.266 and RLVC, which represent the state-of-the-art traditional and end-to-end learned codecs, respectively, and to the top performing end-to-end learned approach in 2021 CLIC competition, E2E_T_OL. Our codec clearly outperforms E2E_T_OL, and compare favorably to VVC and RLVC in some settings.

Vladimir Iashin, Esa Rahtu

Recent advances in visually-induced audio generation are based on sampling short, low-fidelity, and one-class sounds. Moreover, sampling 1 second of audio from the state-of-the-art model takes minutes on a high-end GPU. In this work, we propose a single model capable of generating visually relevant, high-fidelity sounds prompted with a set of frames from open-domain videos in less time than it takes to play it on a single GPU. We train a transformer to sample a new spectrogram from the pre-trained spectrogram codebook given the set of video features. The codebook is obtained using a variant of VQGAN trained to produce a compact sampling space with a novel spectrogram-based perceptual loss. The generated spectrogram is transformed into a waveform using a window-based GAN that significantly speeds up generation. Considering the lack of metrics for automatic evaluation of generated spectrograms, we also build a family of metrics called FID and MKL. These metrics are based on a novel sound classifier, called Melception, and designed to evaluate the fidelity and relevance of open-domain samples. Both qualitative and quantitative studies are conducted on small- and large-scale datasets to evaluate the fidelity and relevance of generated samples. We also compare our model to the state-of-the-art and observe a substantial improvement in quality, size, and computation time. Code, demo, and samples: v-iashin.github.io/SpecVQGAN

Lingyu Zhu, Esa Rahtu

The objective of this paper is to perform visual sound separation: i) we study visual sound separation on spectrograms of different temporal resolutions; ii) we propose a new light yet efficient three-stream framework V-SlowFast that operates on Visual frame, Slow spectrogram, and Fast spectrogram. The Slow spectrogram captures the coarse temporal resolution while the Fast spectrogram contains the fine-grained temporal resolution; iii) we introduce two contrastive objectives to encourage the network to learn discriminative visual features for separating sounds; iv) we propose an audio-visual global attention module for audio and visual feature fusion; v) the introduced V-SlowFast model outperforms previous state-of-the-art in single-frame based visual sound separation on small- and large-scale datasets: MUSIC-21, AVE, and VGG-Sound. We also propose a small V-SlowFast architecture variant, which achieves 74.2% reduction in the number of model parameters and 81.4% reduction in GMACs compared to the previous multi-stage models. Project page: https://ly-zhu.github.io/V-SlowFast

Lam Huynh, Phong Nguyen, Jiri Matas et al.

This paper presents a novel neural architecture search method, called LiDNAS, for generating lightweight monocular depth estimation models. Unlike previous neural architecture search (NAS) approaches, where finding optimized networks are computationally highly demanding, the introduced novel Assisted Tabu Search leads to efficient architecture exploration. Moreover, we construct the search space on a pre-defined backbone network to balance layer diversity and search space size. The LiDNAS method outperforms the state-of-the-art NAS approach, proposed for disparity and depth estimation, in terms of search efficiency and output model performance. The LiDNAS optimized models achieve results superior to compact depth estimation state-of-the-art on NYU-Depth-v2, KITTI, and ScanNet, while being 7%-500% more compact in size, i.e the number of model parameters.

Lam Huynh, Matteo Pedone, Phong Nguyen et al.

Deep neural networks have recently thrived on single image depth estimation. That being said, current developments on this topic highlight an apparent compromise between accuracy and network size. This work proposes an accurate and lightweight framework for monocular depth estimation based on a self-attention mechanism stemming from salient point detection. Specifically, we utilize a sparse set of keypoints to train a FuSaNet model that consists of two major components: Fusion-Net and Saliency-Net. In addition, we introduce a normalized Hessian loss term invariant to scaling and shear along the depth direction, which is shown to substantially improve the accuracy. The proposed method achieves state-of-the-art results on NYU-Depth-v2 and KITTI while using 3.1-38.4 times smaller model in terms of the number of parameters than baseline approaches. Experiments on the SUN-RGBD further demonstrate the generalizability of the proposed method.

Nam Le, Honglei Zhang, Francesco Cricri et al.

Over recent years, deep learning-based computer vision systems have been applied to images at an ever-increasing pace, oftentimes representing the only type of consumption for those images. Given the dramatic explosion in the number of images generated per day, a question arises: how much better would an image codec targeting machine-consumption perform against state-of-the-art codecs targeting human-consumption? In this paper, we propose an image codec for machines which is neural network (NN) based and end-to-end learned. In particular, we propose a set of training strategies that address the delicate problem of balancing competing loss functions, such as computer vision task losses, image distortion losses, and rate loss. Our experimental results show that our NN-based codec outperforms the state-of-the-art Versa-tile Video Coding (VVC) standard on the object detection and instance segmentation tasks, achieving -37.87% and -32.90% of BD-rate gain, respectively, while being fast thanks to its compact size. To the best of our knowledge, this is the first end-to-end learned machine-targeted image codec.

Nam Le, Honglei Zhang, Francesco Cricri et al.

Today, according to the Cisco Annual Internet Report (2018-2023), the fastest-growing category of Internet traffic is machine-to-machine communication. In particular, machine-to-machine communication of images and videos represents a new challenge and opens up new perspectives in the context of data compression. One possible solution approach consists of adapting current human-targeted image and video coding standards to the use case of machine consumption. Another approach consists of developing completely new compression paradigms and architectures for machine-to-machine communications. In this paper, we focus on image compression and present an inference-time content-adaptive finetuning scheme that optimizes the latent representation of an end-to-end learned image codec, aimed at improving the compression efficiency for machine-consumption. The conducted experiments show that our online finetuning brings an average bitrate saving (BD-rate) of -3.66% with respect to our pretrained image codec. In particular, at low bitrate points, our proposed method results in a significant bitrate saving of -9.85%. Overall, our pretrained-and-then-finetuned system achieves -30.54% BD-rate over the state-of-the-art image/video codec Versatile Video Coding (VVC).

Bishwo Adhikari, Esa Rahtu, Heikki Huttunen

Supervised object detection has been proven to be successful in many benchmark datasets achieving human-level performances. However, acquiring a large amount of labeled image samples for supervised detection training is tedious, time-consuming, and costly. In this paper, we propose an efficient image selection approach that samples the most informative images from the unlabeled dataset and utilizes human-machine collaboration in an iterative train-annotate loop. Image features are extracted by the CNN network followed by the similarity score calculation, Euclidean distance. Unlabeled images are then sampled into different approaches based on the similarity score. The proposed approach is straightforward, simple and sampling takes place prior to the network training. Experiments on datasets show that our method can reduce up to 80% of manual annotation workload, compared to full manual labeling setting, and performs better than random sampling.

Saeed Bakhshi Germi, Esa Rahtu, Heikki Huttunen

In this paper, we propose a simple yet effective method to deal with the violation of the Closed-World Assumption for a classifier. Previous works tend to apply a threshold either on the classification scores or the loss function to reject the inputs that violate the assumption. However, these methods cannot achieve the low False Positive Ratio (FPR) required in safety applications. The proposed method is a rejection option based on hypothesis testing with probabilistic networks. With probabilistic networks, it is possible to estimate the distribution of outcomes instead of a single output. By utilizing Z-test over the mean and standard deviation for each class, the proposed method can estimate the statistical significance of the network certainty and reject uncertain outputs. The proposed method was experimented on with different configurations of the COCO and CIFAR datasets. The performance of the proposed method is compared with the Softmax Response, which is a known top-performing method. It is shown that the proposed method can achieve a broader range of operation and cover a lower FPR than the alternative.

Lingyu Zhu, Esa Rahtu

The objective of this paper is to perform audio-visual sound source separation, i.e.~to separate component audios from a mixture based on the videos of sound sources. Moreover, we aim to pinpoint the source location in the input video sequence. Recent works have shown impressive audio-visual separation results when using prior knowledge of the source type (e.g. human playing instrument) and pre-trained motion detectors (e.g. keypoints or optical flows). However, at the same time, the models are limited to a certain application domain. In this paper, we address these limitations and make the following contributions: i) we propose a two-stage architecture, called Appearance and Motion network (AMnet), where the stages specialise to appearance and motion cues, respectively. The entire system is trained in a self-supervised manner; ii) we introduce an Audio-Motion Embedding (AME) framework to explicitly represent the motions that related to sound; iii) we propose an audio-motion transformer architecture for audio and motion feature fusion; iv) we demonstrate state-of-the-art performance on two challenging datasets (MUSIC-21 and AVE) despite the fact that we do not use any pre-trained keypoint detectors or optical flow estimators. Project page: https://ly-zhu.github.io/self-supervised-motion-representations

Soumya Tripathy, Juho Kannala, Esa Rahtu

Image reenactment is a task where the target object in the source image imitates the motion represented in the driving image. One of the most common reenactment tasks is face image animation. The major challenge in the current face reenactment approaches is to distinguish between facial motion and identity. For this reason, the previous models struggle to produce high-quality animations if the driving and source identities are different (cross-person reenactment). We propose a new (face) reenactment model that learns shape-independent motion features in a self-supervised setup. The motion is represented using a set of paired feature points extracted from the source and driving images simultaneously. The model is generalised to multiple reenactment tasks including faces and non-face objects using only a single source image. The extensive experiments show that the model faithfully transfers the driving motion to the source while retaining the source identity intact.

Lam Huynh, Phong Nguyen-Ha, Jiri Matas et al.

In this paper, we propose enhancing monocular depth estimation by adding 3D points as depth guidance. Unlike existing depth completion methods, our approach performs well on extremely sparse and unevenly distributed point clouds, which makes it agnostic to the source of the 3D points. We achieve this by introducing a novel multi-scale 3D point fusion network that is both lightweight and efficient. We demonstrate its versatility on two different depth estimation problems where the 3D points have been acquired with conventional structure-from-motion and LiDAR. In both cases, our network performs on par with state-of-the-art depth completion methods and achieves significantly higher accuracy when only a small number of points is used while being more compact in terms of the number of parameters. We show that our method outperforms some contemporary deep learning based multi-view stereo and structure-from-motion methods both in accuracy and in compactness.