Takuya Narihira

Kazuki Ozeki, Shun Kenney, Yuto Shibata et al.

Although dynamic 3D (i.e., 4D) reconstruction from a monocular dynamic camera has recently advanced, it remains fundamentally limited by depth ambiguity. In this paper, we focus on an alternative practical way, i.e., sparse dynamic camera setup, where a handful of independently moving cameras capture the same subjects. While keeping capture costs low, this setup introduces multi-view constraints and remains practical for real-world video production such as sports, concerts, and TV shows. Despite its potential, our experiments show that naive extensions of existing monocular or dense-fixed camera-based methods are insufficient since they fail to resolve the complex spatiotemporal inconsistencies across views and time. To fill this gap, we propose a simple yet effective 3D track initialization method designed to ensure spatiotemporal consistency by integrating inter-camera feature matching with intra-camera point tracking. Additionally, we incorporate a noise-robust depth-ordering regularization loss and a spatiotemporally diverse batch sampling strategy to enhance optimization stability and cross-view generalization. Furthermore, to address the lack of standardized benchmarks for this task, we introduce LetCamsGo, a new real-world video dataset with 5 sequences across 4 diverse environments, recorded by three independently moving cameras and one fixed camera. Comprehensive benchmarking on LetCamsGo demonstrated that our proposed framework improves 4D reconstruction quality in dynamic regions compared with baselines, paving the way for a low-cost 4D reconstruction paradigm in the wild.

Mungyeom Kim, Minkyeong Jeon, Honggyu An et al.

Dynamic scene reconstruction from monocular video remains a fundamental challenge in computer vision. Existing feed-forward methods predict 3D Gaussians pixel-wise for each frame, suffering from duplicated Gaussians and view-dependent biases that hinder effective learning of scene motion. We present C4G, a feed-forward 4D reconstruction framework built upon a compact set of timestamp-conditioned learnable Gaussian query tokens. Each token aggregates corresponding features across the full temporal context and decodes a 3D Gaussian whose position is modulated by the target timestamp, enabling globally coherent motion modeling without per-scene optimization. To capture fine-grained details, we further introduce a video diffusion model-based rendering enhancement module. Since our framework effectively aggregates features into Gaussians, we extend this capability to feature lifting, producing a 4D feature field that supports point tracking and dynamic scene understanding. C4G achieves strong novel-view synthesis performance using significantly fewer Gaussians and without requiring camera poses, while exhibiting stronger motion modeling and robustness to large temporal gaps.

Hiromichi Kamata, Yuiko Sakuma, Akio Hayakawa et al.

We propose a high-quality 3D-to-3D conversion method, Instruct 3D-to-3D. Our method is designed for a novel task, which is to convert a given 3D scene to another scene according to text instructions. Instruct 3D-to-3D applies pretrained Image-to-Image diffusion models for 3D-to-3D conversion. This enables the likelihood maximization of each viewpoint image and high-quality 3D generation. In addition, our proposed method explicitly inputs the source 3D scene as a condition, which enhances 3D consistency and controllability of how much of the source 3D scene structure is reflected. We also propose dynamic scaling, which allows the intensity of the geometry transformation to be adjusted. We performed quantitative and qualitative evaluations and showed that our proposed method achieves higher quality 3D-to-3D conversions than baseline methods.

Naoki Matsunaga, Masato Ishii, Akio Hayakawa et al.

Our goal is to develop fine-grained real-image editing methods suitable for real-world applications. In this paper, we first summarize four requirements for these methods and propose a novel diffusion-based image editing framework with pixel-wise guidance that satisfies these requirements. Specifically, we train pixel-classifiers with a few annotated data and then infer the segmentation map of a target image. Users then manipulate the map to instruct how the image will be edited. We utilize a pre-trained diffusion model to generate edited images aligned with the user's intention with pixel-wise guidance. The effective combination of proposed guidance and other techniques enables highly controllable editing with preserving the outside of the edited area, which results in meeting our requirements. The experimental results demonstrate that our proposal outperforms the GAN-based method for editing quality and speed.

Yuiko Sakuma, Masato Ishii, Takuya Narihira

We address the challenge of training a large supernet for the object detection task, using a relatively small amount of training data. Specifically, we propose an efficient supernet-based neural architecture search (NAS) method that uses search space pruning. The search space defined by the supernet is pruned by removing candidate models that are predicted to perform poorly. To effectively remove the candidates over a wide range of resource constraints, we particularly design a performance predictor for supernet, called path filter, which is conditioned by resource constraints and can accurately predict the relative performance of the models that satisfy similar resource constraints. Hence, supernet training is more focused on the best-performing candidates. Our path filter handles prediction for paths with different resource budgets. Compared to once-for-all, our proposed method reduces the computational cost of the optimal network architecture by 30% and 63%, while yielding better accuracy-floating point operations Pareto front (0.85 and 0.45 points of improvement on average precision for Pascal VOC and COCO, respectively).

Kazuki Yoshiyama, Takuya Narihira

The goal of inverse rendering is to decompose geometry, lights, and materials given pose multi-view images. To achieve this goal, we propose neural direct and joint inverse rendering, NDJIR. Different from prior works which relies on some approximations of the rendering equation, NDJIR directly addresses the integrals in the rendering equation and jointly decomposes geometry: signed distance function, lights: environment and implicit lights, materials: base color, roughness, specular reflectance using the powerful and flexible volume rendering framework, voxel grid feature, and Bayesian prior. Our method directly uses the physically-based rendering, so we can seamlessly export an extracted mesh with materials to DCC tools and show material conversion examples. We perform intensive experiments to show that our proposed method can decompose semantically well for real object in photogrammetric setting and what factors contribute towards accurate inverse rendering.

Honggyu An, Jaewoo Jung, Mungyeom Kim et al.

Reconstructing and understanding 3D scenes from unposed sparse views in a feed-forward manner remains as a challenging task in 3D computer vision. Recent approaches use per-pixel 3D Gaussian Splatting for reconstruction, followed by a 2D-to-3D feature lifting stage for scene understanding. However, they generate excessive redundant Gaussians, causing high memory overhead and sub-optimal multi-view feature aggregation, leading to degraded novel view synthesis and scene understanding performance. We propose C3G, a novel feed-forward framework that estimates compact 3D Gaussians only at essential spatial locations, minimizing redundancy while enabling effective feature lifting. We introduce learnable tokens that aggregate multi-view features through self-attention to guide Gaussian generation, ensuring each Gaussian integrates relevant visual features across views. We then exploit the learned attention patterns for Gaussian decoding to efficiently lift features. Extensive experiments on pose-free novel view synthesis, 3D open-vocabulary segmentation, and view-invariant feature aggregation demonstrate our approach's effectiveness. Results show that a compact yet geometrically meaningful representation is sufficient for high-quality scene reconstruction and understanding, achieving superior memory efficiency and feature fidelity compared to existing methods.

Jisang Han, Honggyu An, Jaewoo Jung et al.

In this work, we address the task of 3D reconstruction in dynamic scenes, where object motions frequently degrade the quality of previous 3D pointmap regression methods, such as DUSt3R, that are originally designed for static 3D scene reconstruction. Although these methods provide an elegant and powerful solution in static settings, they struggle in the presence of dynamic motions that disrupt alignment based solely on camera poses. To overcome this, we propose $D^2USt3R$ that directly regresses Static-Dynamic Aligned Pointmaps (SDAP) that simultaneiously capture both static and dynamic 3D scene geometry. By explicitly incorporating both spatial and temporal aspects, our approach successfully encapsulates 3D dense correspondence to the proposed pointmaps, enhancing downstream tasks. Extensive experimental evaluations demonstrate that our proposed approach consistently achieves superior 3D reconstruction performance across various datasets featuring complex motions.

Junyoung Seo, Jisang Han, Jaewoo Jung et al.

We introduce Vid-CamEdit, a novel framework for video camera trajectory editing, enabling the re-synthesis of monocular videos along user-defined camera paths. This task is challenging due to its ill-posed nature and the limited multi-view video data for training. Traditional reconstruction methods struggle with extreme trajectory changes, and existing generative models for dynamic novel view synthesis cannot handle in-the-wild videos. Our approach consists of two steps: estimating temporally consistent geometry, and generative rendering guided by this geometry. By integrating geometric priors, the generative model focuses on synthesizing realistic details where the estimated geometry is uncertain. We eliminate the need for extensive 4D training data through a factorized fine-tuning framework that separately trains spatial and temporal components using multi-view image and video data. Our method outperforms baselines in producing plausible videos from novel camera trajectories, especially in extreme extrapolation scenarios on real-world footage.

Shoukang Hu, Takuya Narihira, Kazumi Fukuda et al.

Previous 3D human creation methods have made significant progress in synthesizing view-consistent and temporally aligned results from sparse-view images or monocular videos. However, it remains challenging to produce perpetually realistic, view-consistent, and temporally coherent human avatars from a single image, as limited information is available in the single-view input setting. Motivated by the success of 2D character animation, we propose HumanGif, a single-view human diffusion model with generative prior. Specifically, we formulate the single-view-based 3D human novel view and pose synthesis as a single-view-conditioned human diffusion process, utilizing generative priors from foundational diffusion models to complement the missing information. To ensure fine-grained and consistent novel view and pose synthesis, we introduce a Human NeRF module in HumanGif to learn spatially aligned features from the input image, implicitly capturing the relative camera and human pose transformation. Furthermore, we introduce an image-level loss during optimization to bridge the gap between latent and image spaces in diffusion models. Extensive experiments on RenderPeople, DNA-Rendering, THuman 2.1, and TikTok datasets demonstrate that HumanGif achieves the best perceptual performance, with better generalizability for novel view and pose synthesis.

JoungBin Lee, Jaewoo Jung, Jisang Han et al.

We present 3DScenePrompt, a framework that generates the next video chunk from arbitrary-length input while enabling precise camera control and preserving scene consistency. Unlike methods conditioned on a single image or a short clip, we employ dual spatio-temporal conditioning that reformulates context-view referencing across the input video. Our approach conditions on both temporally adjacent frames for motion continuity and spatially adjacent content for scene consistency. However, when generating beyond temporal boundaries, directly using spatially adjacent frames would incorrectly preserve dynamic elements from the past. We address this by introducing a 3D scene memory that represents exclusively the static geometry extracted from the entire input video. To construct this memory, we leverage dynamic SLAM with our newly introduced dynamic masking strategy that explicitly separates static scene geometry from moving elements. The static scene representation can then be projected to any target viewpoint, providing geometrically consistent warped views that serve as strong 3D spatial prompts while allowing dynamic regions to evolve naturally from temporal context. This enables our model to maintain long-range spatial coherence and precise camera control without sacrificing computational efficiency or motion realism. Extensive experiments demonstrate that our framework significantly outperforms existing methods in scene consistency, camera controllability, and generation quality. Project page : https://cvlab-kaist.github.io/3DScenePrompt/

Takuya Yashima, Takuya Narihira, Tamaki Kojima

In recent advances of deep generative models, face reenactment -manipulating and controlling human face, including their head movement-has drawn much attention for its wide range of applicability. Despite its strong expressiveness, it is inevitable that the models fail to reconstruct or accurately generate unseen side of the face of a given single reference image. Most of existing methods alleviate this problem by learning appearances of human faces from large amount of data and generate realistic texture at inference time. Rather than completely relying on what generative models learn, we show that simple extension by using multiple reference images significantly improves generation quality. We show this by 1) conducting the reconstruction task on publicly available dataset, 2) conducting facial motion transfer on our original dataset which consists of multi-person's head movement video sequences, and 3) using a newly proposed evaluation metric to validate that our method achieves better quantitative results.

Kenji Suzuki, Yoshiyuki Kobayashi, Takuya Narihira

Identifying the influence of training data for data cleansing can improve the accuracy of deep learning. An approach with stochastic gradient descent (SGD) called SGD-influence to calculate the influence scores was proposed, but, the calculation costs are expensive. It is necessary to temporally store the parameters of the model during training phase for inference phase to calculate influence sores. In close connection with the previous method, we propose a method to reduce cache files to store the parameters in training phase for calculating inference score. We only adopt the final parameters in last epoch for influence functions calculation. In our experiments on classification, the cache size of training using MNIST dataset with our approach is 1.236 MB. On the other hand, the previous method used cache size of 1.932 GB in last epoch. It means that cache size has been reduced to 1/1,563. We also observed the accuracy improvement by data cleansing with removal of negatively influential data using our approach as well as the previous method. Moreover, our simple and general proposed method to calculate influence scores is available on our auto ML tool without programing, Neural Network Console. The source code is also available.

Andrew Shin, Masato Ishii, Takuya Narihira

Transformer architectures have brought about fundamental changes to computational linguistic field, which had been dominated by recurrent neural networks for many years. Its success also implies drastic changes in cross-modal tasks with language and vision, and many researchers have already tackled the issue. In this paper, we review some of the most critical milestones in the field, as well as overall trends on how transformer architecture has been incorporated into visuolinguistic cross-modal tasks. Furthermore, we discuss its current limitations and speculate upon some of the prospects that we find imminent.

Takuya Narihira, Javier Alonsogarcia, Fabien Cardinaux et al.

While there exist a plethora of deep learning tools and frameworks, the fast-growing complexity of the field brings new demands and challenges, such as more flexible network design, speedy computation on distributed setting, and compatibility between different tools. In this paper, we introduce Neural Network Libraries (https://nnabla.org), a deep learning framework designed from engineer's perspective, with emphasis on usability and compatibility as its core design principles. We elaborate on each of our design principles and its merits, and validate our attempts via experiments.

Naofumi Akimoto, Akio Hayakawa, Andrew Shin et al.

We propose a novel reference-based video colorization framework with spatiotemporal correspondence. Reference-based methods colorize grayscale frames referencing a user input color frame. Existing methods suffer from the color leakage between objects and the emergence of average colors, derived from non-local semantic correspondence in space. To address this issue, we warp colors only from the regions on the reference frame restricted by correspondence in time. We propagate masks as temporal correspondences, using two complementary tracking approaches: off-the-shelf instance tracking for high performance segmentation, and newly proposed dense tracking to track various types of objects. By restricting temporally-related regions for referencing colors, our approach propagates faithful colors throughout the video. Experiments demonstrate that our method outperforms state-of-the-art methods quantitatively and qualitatively.

Akio Hayakawa, Takuya Narihira

While large neural networks demonstrate higher performance in various tasks, training large networks is difficult due to limitations on GPU memory size. We propose a novel out-of-core algorithm that enables faster training of extremely large-scale neural networks with sizes larger than allotted GPU memory. Under a given memory budget constraint, our scheduling algorithm locally adapts the timing of memory transfers according to memory usage of each function, which improves overlap between computation and memory transfers. Additionally, we apply virtual addressing technique, commonly performed in OS, to training of neural networks with out-of-core execution, which drastically reduces the amount of memory fragmentation caused by frequent memory transfers. With our proposed algorithm, we successfully train ResNet-50 with 1440 batch-size with keeping training speed at 55%, which is 7.5x larger than the upper bound of physical memory. It also outperforms a previous state-of-the-art substantially, i.e. it trains a 1.55x larger network than state-of-the-art with faster execution. Moreover, we experimentally show that our approach is also scalable for various types of networks.

Yuka Ariki, Takuya Narihira

Path-planning algorithms are an important part of a wide variety of robotic applications, such as mobile robot navigation and robot arm manipulation. However, in large search spaces in which local traps may exist, it remains challenging to reliably find a path while satisfying real-time constraints. Efforts to speed up the path search have led to the development of many practical path-planning algorithms. These algorithms often define a search heuristic to guide the search towards the goal. The heuristics should be carefully designed for each specific problem to ensure reliability in the various situations encountered in the problem. However, it is often difficult for humans to craft such robust heuristics, and the search performance often degrades under conditions that violate the heuristic assumption. Rather than manually designing the heuristics, in this work, we propose a learning approach to acquire these search heuristics. Our method represents the environment containing the obstacles as an image, and this image is fed into fully convolutional neural networks to produce a search heuristic image where every pixel represents a heuristic value (cost-to-go value to a goal) in the form of a vertex of a search graph. Training the heuristic is performed using previously collected planning results. Our preliminary experiments (2D grid world navigation experiments) demonstrate significant reduction in the search costs relative to a hand-designed heuristic.

Michael Maire, Takuya Narihira, Stella X. Yu

Spectral embedding provides a framework for solving perceptual organization problems, including image segmentation and figure/ground organization. From an affinity matrix describing pairwise relationships between pixels, it clusters pixels into regions, and, using a complex-valued extension, orders pixels according to layer. We train a convolutional neural network (CNN) to directly predict the pairwise relationships that define this affinity matrix. Spectral embedding then resolves these predictions into a globally-consistent segmentation and figure/ground organization of the scene. Experiments demonstrate significant benefit to this direct coupling compared to prior works which use explicit intermediate stages, such as edge detection, on the pathway from image to affinities. Our results suggest spectral embedding as a powerful alternative to the conditional random field (CRF)-based globalization schemes typically coupled to deep neural networks.

Takuya Narihira, Michael Maire, Stella X. Yu

We introduce a new approach to intrinsic image decomposition, the task of decomposing a single image into albedo and shading components. Our strategy, which we term direct intrinsics, is to learn a convolutional neural network (CNN) that directly predicts output albedo and shading channels from an input RGB image patch. Direct intrinsics is a departure from classical techniques for intrinsic image decomposition, which typically rely on physically-motivated priors and graph-based inference algorithms. The large-scale synthetic ground-truth of the MPI Sintel dataset plays a key role in training direct intrinsics. We demonstrate results on both the synthetic images of Sintel and the real images of the classic MIT intrinsic image dataset. On Sintel, direct intrinsics, using only RGB input, outperforms all prior work, including methods that rely on RGB+Depth input. Direct intrinsics also generalizes across modalities; it produces quite reasonable decompositions on the real images of the MIT dataset. Our results indicate that the marriage of CNNs with synthetic training data may be a powerful new technique for tackling classic problems in computer vision.

Takuya Narihira, Damian Borth, Stella X. Yu et al.

We consider the visual sentiment task of mapping an image to an adjective noun pair (ANP) such as "cute baby". To capture the two-factor structure of our ANP semantics as well as to overcome annotation noise and ambiguity, we propose a novel factorized CNN model which learns separate representations for adjectives and nouns but optimizes the classification performance over their product. Our experiments on the publicly available SentiBank dataset show that our model significantly outperforms not only independent ANP classifiers on unseen ANPs and on retrieving images of novel ANPs, but also image captioning models which capture word semantics from co-occurrence of natural text; the latter turn out to be surprisingly poor at capturing the sentiment evoked by pure visual experience. That is, our factorized ANP CNN not only trains better from noisy labels, generalizes better to new images, but can also expands the ANP vocabulary on its own.