Cheng Sun

Tao Tu, Shun-Po Chuang, Yu-Lun Liu et al. · nvidia

We propose ImGeoNet, a multi-view image-based 3D object detection framework that models a 3D space by an image-induced geometry-aware voxel representation. Unlike previous methods which aggregate 2D features into 3D voxels without considering geometry, ImGeoNet learns to induce geometry from multi-view images to alleviate the confusion arising from voxels of free space, and during the inference phase, only images from multiple views are required. Besides, a powerful pre-trained 2D feature extractor can be leveraged by our representation, leading to a more robust performance. To evaluate the effectiveness of ImGeoNet, we conduct quantitative and qualitative experiments on three indoor datasets, namely ARKitScenes, ScanNetV2, and ScanNet200. The results demonstrate that ImGeoNet outperforms the current state-of-the-art multi-view image-based method, ImVoxelNet, on all three datasets in terms of detection accuracy. In addition, ImGeoNet shows great data efficiency by achieving results comparable to ImVoxelNet with 100 views while utilizing only 40 views. Furthermore, our studies indicate that our proposed image-induced geometry-aware representation can enable image-based methods to attain superior detection accuracy than the seminal point cloud-based method, VoteNet, in two practical scenarios: (1) scenarios where point clouds are sparse and noisy, such as in ARKitScenes, and (2) scenarios involve diverse object classes, particularly classes of small objects, as in the case in ScanNet200.

Cheng Sun, Guangyan Cai, Zhengqin Li et al. · nvidia

Reconstructing the shape and spatially varying surface appearances of a physical-world object as well as its surrounding illumination based on 2D images (e.g., photographs) of the object has been a long-standing problem in computer vision and graphics. In this paper, we introduce an accurate and highly efficient object reconstruction pipeline combining neural based object reconstruction and physics-based inverse rendering (PBIR). Our pipeline firstly leverages a neural SDF based shape reconstruction to produce high-quality but potentially imperfect object shape. Then, we introduce a neural material and lighting distillation stage to achieve high-quality predictions for material and illumination. In the last stage, initialized by the neural predictions, we perform PBIR to refine the initial results and obtain the final high-quality reconstruction of object shape, material, and illumination. Experimental results demonstrate our pipeline significantly outperforms existing methods quality-wise and performance-wise.

Yi-Chia Chen, Wei-Hua Li, Cheng Sun et al. · nvidia

We introduce SAM4MLLM, an innovative approach which integrates the Segment Anything Model (SAM) with Multi-Modal Large Language Models (MLLMs) for pixel-aware tasks. Our method enables MLLMs to learn pixel-level location information without requiring excessive modifications to the existing model architecture or adding specialized tokens. We introduce an inquiry-based approach that can effectively find prompt points for SAM to perform segmentation based on MLLM. It combines detailed visual information with the powerful expressive capabilities of large language models in a unified language-based manner without additional computational overhead in learning. Experimental results on pubic benchmarks demonstrate the effectiveness of our approach.

Cheng Sun, Wei-En Tai, Yu-Lin Shih et al. · nvidia

State-of-the-art single-view 360-degree room layout reconstruction methods formulate the problem as a high-level 1D (per-column) regression task. On the other hand, traditional low-level 2D layout segmentation is simpler to learn and can represent occluded regions, but it requires complex post-processing for the targeting layout polygon and sacrifices accuracy. We present Seg2Reg to render 1D layout depth regression from the 2D segmentation map in a differentiable and occlusion-aware way, marrying the merits of both sides. Specifically, our model predicts floor-plan density for the input equirectangular 360-degree image. Formulating the 2D layout representation as a density field enables us to employ `flattened' volume rendering to form 1D layout depth regression. In addition, we propose a novel 3D warping augmentation on layout to improve generalization. Finally, we re-implement recent room layout reconstruction methods into our codebase for benchmarking and explore modern backbones and training techniques to serve as the strong baseline. Our model significantly outperforms previous arts. The code will be made available upon publication.

Cheng-Hung Chan, Cheng-Yang Yuan, Cheng Sun et al. · nvidia

We present a video decomposition method that facilitates layer-based editing of videos with spatiotemporally varying lighting and motion effects. Our neural model decomposes an input video into multiple layered representations, each comprising a 2D texture map, a mask for the original video, and a multiplicative residual characterizing the spatiotemporal variations in lighting conditions. A single edit on the texture maps can be propagated to the corresponding locations in the entire video frames while preserving other contents' consistencies. Our method efficiently learns the layer-based neural representations of a 1080p video in 25s per frame via coordinate hashing and allows real-time rendering of the edited result at 71 fps on a single GPU. Qualitatively, we run our method on various videos to show its effectiveness in generating high-quality editing effects. Quantitatively, we propose to adopt feature-tracking evaluation metrics for objectively assessing the consistency of video editing. Project page: https://lightbulb12294.github.io/hashing-nvd/

Chih-Jung Tsai, Cheng Sun, Hwann-Tzong Chen · nvidia

This paper aims to address a new task of image morphing under a multiview setting, which takes two sets of multiview images as the input and generates intermediate renderings that not only exhibit smooth transitions between the two input sets but also ensure visual consistency across different views at any transition state. To achieve this goal, we propose a novel approach called Multiview Regenerative Morphing that formulates the morphing process as an optimization to solve for rigid transformation and optimal-transport interpolation. Given the multiview input images of the source and target scenes, we first learn a volumetric representation that models the geometry and appearance for each scene to enable the rendering of novel views. Then, the morphing between the two scenes is obtained by solving optimal transport between the two volumetric representations in Wasserstein metrics. Our approach does not rely on user-specified correspondences or 2D/3D input meshes, and we do not assume any predefined categories of the source and target scenes. The proposed view-consistent interpolation scheme directly works on multiview images to yield a novel and visually plausible effect of multiview free-form morphing.

Ping-Chung Yu, Cheng Sun, Min Sun · nvidia

Collecting and labeling the registered 3D point cloud is costly. As a result, 3D resources for training are typically limited in quantity compared to the 2D images counterpart. In this work, we deal with the data scarcity challenge of 3D tasks by transferring knowledge from strong 2D models via RGB-D images. Specifically, we utilize a strong and well-trained semantic segmentation model for 2D images to augment RGB-D images with pseudo-label. The augmented dataset can then be used to pre-train 3D models. Finally, by simply fine-tuning on a few labeled 3D instances, our method already outperforms existing state-of-the-art that is tailored for 3D label efficiency. We also show that the results of mean-teacher and entropy minimization can be improved by our pre-training, suggesting that the transferred knowledge is helpful in semi-supervised setting. We verify the effectiveness of our approach on two popular 3D models and three different tasks. On ScanNet official evaluation, we establish new state-of-the-art semantic segmentation results on the data-efficient track.

Cheng Sun, Jaesung Choe, Min-Hung Chen et al.

Recent Large View Synthesis Models (LVSMs) advocate an encoder-decoder architecture that separates reconstruction and rendering into distinct networks. We re-examine this design. Through controlled experiments, we show that a decoder-only architecture, which represents scenes implicitly as a KV-cache, outperforms encoder-decoder variants while using fewer parameters at identical rendering complexity. Further analysis shows that sharing weights between the color-input reconstruction network and the camera-only rendering network better aligns their features at the same viewpoint, facilitating image synthesis. Building on this finding, our model, dubbed DVSM, further incorporates foundation model priors and stage-wise patch sizing for an improved efficiency-quality tradeoff. Our results establish a new state of the art for novel-view synthesis across multiple benchmarks, in some cases even outperforming per-scene-optimized 3DGS under dense input views.

Yu-Cheng Hsieh, Cheng Sun, Suraj Dengale et al. · nvidia

The volume and diversity of training data are critical for modern deep learningbased methods. Compared to the massive amount of labeled perspective images, 360 panoramic images fall short in both volume and diversity. In this paper, we propose PanoMixSwap, a novel data augmentation technique specifically designed for indoor panoramic images. PanoMixSwap explicitly mixes various background styles, foreground furniture, and room layouts from the existing indoor panorama datasets and generates a diverse set of new panoramic images to enrich the datasets. We first decompose each panoramic image into its constituent parts: background style, foreground furniture, and room layout. Then, we generate an augmented image by mixing these three parts from three different images, such as the foreground furniture from one image, the background style from another image, and the room structure from the third image. Our method yields high diversity since there is a cubical increase in image combinations. We also evaluate the effectiveness of PanoMixSwap on two indoor scene understanding tasks: semantic segmentation and layout estimation. Our experiments demonstrate that state-of-the-art methods trained with PanoMixSwap outperform their original setting on both tasks consistently.

Meituan LongCat Team, Bayan, Bei Li et al.

We introduce LongCat-Flash, a 560-billion-parameter Mixture-of-Experts (MoE) language model designed for both computational efficiency and advanced agentic capabilities. Stemming from the need for scalable efficiency, LongCat-Flash adopts two novel designs: (a) Zero-computation Experts, which enables dynamic computational budget allocation and activates 18.6B-31.3B (27B on average) per token depending on contextual demands, optimizing resource usage. (b) Shortcut-connected MoE, which enlarges the computation-communication overlap window, demonstrating notable gains in inference efficiency and throughput compared to models of a comparable scale. We develop a comprehensive scaling framework for large models that combines hyperparameter transfer, model-growth initialization, a multi-pronged stability suite, and deterministic computation to achieve stable and reproducible training. Notably, leveraging the synergy among scalable architectural design and infrastructure efforts, we complete model training on more than 20 trillion tokens within 30 days, while achieving over 100 tokens per second (TPS) for inference at a cost of \$0.70 per million output tokens. To cultivate LongCat-Flash towards agentic intelligence, we conduct a large-scale pre-training on optimized mixtures, followed by targeted mid- and post-training on reasoning, code, and instructions, with further augmentation from synthetic data and tool use tasks. Comprehensive evaluations demonstrate that, as a non-thinking foundation model, LongCat-Flash delivers highly competitive performance among other leading models, with exceptional strengths in agentic tasks. The model checkpoint of LongCat-Flash is open-sourced to foster community research. LongCat Chat: https://longcat.ai Hugging Face: https://huggingface.co/meituan-longcat GitHub: https://github.com/meituan-longcat

Zilong Bai, Ruiji Zhang, Linqing Chen et al.

In recent years, large language models(LLMs) have attracted significant attention due to their exceptional performance across a multitude of natural language process tasks, and have been widely applied in various fields. However, the application of large language models in the Intellectual Property (IP) domain is challenging due to the strong need for specialized knowledge, privacy protection, processing of extremely long text in this field. In this technical report, we present for the first time a low-cost, standardized procedure for training IP-oriented LLMs, meeting the unique requirements of the IP domain. Using this standard process, we have trained the PatentGPT series models based on open-source pretrained models. By evaluating them on the open-source IP-oriented benchmark MOZIP, our domain-specific LLMs outperforms GPT-4, indicating the effectiveness of the proposed training procedure and the expertise of the PatentGPT models in the IP domain. Remarkably, our model surpassed GPT-4 on the 2019 China Patent Agent Qualification Examination, scoring 65 and matching human expert levels. Additionally, the PatentGPT model, which utilizes the SMoE architecture, achieves performance comparable to that of GPT-4 in the IP domain and demonstrates a better cost-performance ratio on long-text tasks, potentially serving as an alternative to GPT-4 within the IP domain.

Wei-Hua Li, Cheng Sun, Chu-Song Chen

Multimodal story customization aims to generate coherent story flows conditioned on textual descriptions, reference identity images, and shot types. While recent progress in story generation has shown promising results, most approaches rely on text-only inputs. A few studies incorporate character identity cues (e.g., facial ID), but lack broader multimodal conditioning. In this work, we introduce VstoryGen, a multimodal framework that integrates descriptions with character and background references to enable customizable story generation. To enhance cinematic diversity, we introduce shot-type control via parameter-efficient prompt tuning on movie data, enabling the model to generate sequences that more faithfully reflect cinematic grammar. To evaluate our framework, we establish two new benchmarks that assess multimodal story customization from the perspectives of character and scene consistency, text-visual alignment, and shot-type control. Experiments demonstrate that VstoryGen achieves improved consistency and cinematic diversity compared to existing methods.

Yoonwoo Jeong, Cheng Sun, Frank Wang et al.

Recent advancements in computer vision have successfully extended Open-vocabulary segmentation (OVS) to the 3D domain by leveraging 3D Gaussian Splatting (3D-GS). Despite this progress, efficiently rendering the high-dimensional features required for open-vocabulary queries poses a significant challenge. Existing methods employ codebooks or feature compression, causing information loss, thereby degrading segmentation quality. To address this limitation, we introduce Quantile Rendering (Q-Render), a novel rendering strategy for 3D Gaussians that efficiently handles high-dimensional features while maintaining high fidelity. Unlike conventional volume rendering, which densely samples all 3D Gaussians intersecting each ray, Q-Render sparsely samples only those with dominant influence along the ray. By integrating Q-Render into a generalizable 3D neural network, we also propose Gaussian Splatting Network (GS-Net), which predicts Gaussian features in a generalizable manner. Extensive experiments on ScanNet and LeRF demonstrate that our framework outperforms state-of-the-art methods, while enabling real-time rendering with an approximate ~43.7x speedup on 512-D feature maps. Code will be made publicly available.

Yang-Che Sun, Cheng Sun, Chin-Yang Lin et al.

Video object segmentation methods like SAM2 achieve strong performance through memory-based architectures but struggle under large viewpoint changes due to reliance on appearance features. Traditional 3D instance segmentation methods address viewpoint consistency but require camera poses, depth maps, and expensive preprocessing. We introduce 3AM, a training-time enhancement that integrates 3D-aware features from MUSt3R into SAM2. Our lightweight Feature Merger fuses multi-level MUSt3R features that encode implicit geometric correspondence. Combined with SAM2's appearance features, the model achieves geometry-consistent recognition grounded in both spatial position and visual similarity. We propose a field-of-view aware sampling strategy ensuring frames observe spatially consistent object regions for reliable 3D correspondence learning. Critically, our method requires only RGB input at inference, with no camera poses or preprocessing. On challenging datasets with wide-baseline motion (ScanNet++, Replica), 3AM substantially outperforms SAM2 and extensions, achieving 90.6% IoU and 71.7% Positive IoU on ScanNet++'s Selected Subset, improving over state-of-the-art VOS methods by +15.9 and +30.4 points. Project page: https://jayisaking.github.io/3AM-Page/

Sheng-Yu Huang, Jaesung Choe, Yu-Chiang Frank Wang et al.

We propose OpenVoxel, a training-free algorithm for grouping and captioning sparse voxels for the open-vocabulary 3D scene understanding tasks. Given the sparse voxel rasterization (SVR) model obtained from multi-view images of a 3D scene, our OpenVoxel is able to produce meaningful groups that describe different objects in the scene. Also, by leveraging powerful Vision Language Models (VLMs) and Multi-modal Large Language Models (MLLMs), our OpenVoxel successfully build an informative scene map by captioning each group, enabling further 3D scene understanding tasks such as open-vocabulary segmentation (OVS) or referring expression segmentation (RES). Unlike previous methods, our method is training-free and does not introduce embeddings from a CLIP/BERT text encoder. Instead, we directly proceed with text-to-text search using MLLMs. Through extensive experiments, our method demonstrates superior performance compared to recent studies, particularly in complex referring expression segmentation (RES) tasks. The code will be open.

Hao-Wen Ting, Cheng Sun, Hwann-Tzong Chen

We present the first self-supervised method to train panoramic room layout estimation models without any labeled data. Unlike per-pixel dense depth that provides abundant correspondence constraints, layout representation is sparse and topological, hindering the use of self-supervised reprojection consistency on images. To address this issue, we propose Differentiable Layout View Rendering, which can warp a source image to the target camera pose given the estimated layout from the target image. As each rendered pixel is differentiable with respect to the estimated layout, we can now train the layout estimation model by minimizing reprojection loss. Besides, we introduce regularization losses to encourage Manhattan alignment, ceiling-floor alignment, cycle consistency, and layout stretch consistency, which further improve our predictions. Finally, we present the first self-supervised results on ZilloIndoor and MatterportLayout datasets. Our approach also shows promising solutions in data-scarce scenarios and active learning, which would have an immediate value in the real estate virtual tour software. Code is available at https://github.com/joshua049/Stereo-360-Layout.

Cheng Sun, Jaesung Choe, Charles Loop et al. · nvidia

We propose an efficient radiance field rendering algorithm that incorporates a rasterization process on adaptive sparse voxels without neural networks or 3D Gaussians. There are two key contributions coupled with the proposed system. The first is to adaptively and explicitly allocate sparse voxels to different levels of detail within scenes, faithfully reproducing scene details with $65536^3$ grid resolution while achieving high rendering frame rates. Second, we customize a rasterizer for efficient adaptive sparse voxels rendering. We render voxels in the correct depth order by using ray direction-dependent Morton ordering, which avoids the well-known popping artifact found in Gaussian splatting. Our method improves the previous neural-free voxel model by over 4db PSNR and more than 10x FPS speedup, achieving state-of-the-art comparable novel-view synthesis results. Additionally, our voxel representation is seamlessly compatible with grid-based 3D processing techniques such as Volume Fusion, Voxel Pooling, and Marching Cubes, enabling a wide range of future extensions and applications.

Chin-Yang Lin, Chung-Ho Wu, Chang-Han Yeh et al. · nvidia

Neural Radiance Fields (NeRF) face significant challenges in extreme few-shot scenarios, primarily due to overfitting and long training times. Existing methods, such as FreeNeRF and SparseNeRF, use frequency regularization or pre-trained priors but struggle with complex scheduling and bias. We introduce FrugalNeRF, a novel few-shot NeRF framework that leverages weight-sharing voxels across multiple scales to efficiently represent scene details. Our key contribution is a cross-scale geometric adaptation scheme that selects pseudo ground truth depth based on reprojection errors across scales. This guides training without relying on externally learned priors, enabling full utilization of the training data. It can also integrate pre-trained priors, enhancing quality without slowing convergence. Experiments on LLFF, DTU, and RealEstate-10K show that FrugalNeRF outperforms other few-shot NeRF methods while significantly reducing training time, making it a practical solution for efficient and accurate 3D scene reconstruction.

Chin-Yang Lin, Cheng Sun, Fu-En Yang et al. · nvidia

LongSplat addresses critical challenges in novel view synthesis (NVS) from casually captured long videos characterized by irregular camera motion, unknown camera poses, and expansive scenes. Current methods often suffer from pose drift, inaccurate geometry initialization, and severe memory limitations. To address these issues, we introduce LongSplat, a robust unposed 3D Gaussian Splatting framework featuring: (1) Incremental Joint Optimization that concurrently optimizes camera poses and 3D Gaussians to avoid local minima and ensure global consistency; (2) a robust Pose Estimation Module leveraging learned 3D priors; and (3) an efficient Octree Anchor Formation mechanism that converts dense point clouds into anchors based on spatial density. Extensive experiments on challenging benchmarks demonstrate that LongSplat achieves state-of-the-art results, substantially improving rendering quality, pose accuracy, and computational efficiency compared to prior approaches. Project page: https://linjohnss.github.io/longsplat/

Yu-Heng Hung, Kai-Jie Lin, Yu-Heng Lin et al. · nvidia

Bayesian optimization (BO) offers an efficient pipeline for optimizing black-box functions with the help of a Gaussian process prior and an acquisition function (AF). Recently, in the context of single-objective BO, learning-based AFs witnessed promising empirical results given its favorable non-myopic nature. Despite this, the direct extension of these approaches to multi-objective Bayesian optimization (MOBO) suffer from the \textit{hypervolume identifiability issue}, which results from the non-Markovian nature of MOBO problems. To tackle this, inspired by the non-Markovian RL literature and the success of Transformers in language modeling, we present a generalized deep Q-learning framework and propose \textit{BOFormer}, which substantiates this framework for MOBO via sequence modeling. Through extensive evaluation, we demonstrate that BOFormer constantly outperforms the benchmark rule-based and learning-based algorithms in various synthetic MOBO and real-world multi-objective hyperparameter optimization problems. We have made the source code publicly available to encourage further research in this direction.

Linqing Chen, Weilei Wang, Yubin Xia et al.

In this paper, we propose a novel system that integrates state-of-the-art, domain-specific large language models with advanced information retrieval techniques to deliver comprehensive and context-aware responses. Our approach facilitates seamless interaction among diverse components, enabling cross-validation of outputs to produce accurate, high-quality responses enriched with relevant data, images, tables, and other modalities. We demonstrate the system's capability to enhance response precision by leveraging a robust question-answering model, significantly improving the quality of dialogue generation. The system provides an accessible platform for real-time, high-fidelity interactions, allowing users to benefit from efficient human-computer interaction, precise retrieval, and simultaneous access to a wide range of literature and data. This dramatically improves the research efficiency of professionals in the biomedical and pharmaceutical domains and facilitates faster, more informed decision-making throughout the R\&D process. Furthermore, the system proposed in this paper is available at https://synapse-chat.patsnap.com.

Yoonwoo Jeong, Cheng Sun, Yu-Chiang Frank Wang et al.

Promptable segmentation has emerged as a powerful paradigm in computer vision, enabling users to guide models in parsing complex scenes with prompts such as clicks, boxes, or textual cues. Recent advances, exemplified by the Segment Anything Model (SAM), have extended this paradigm to videos and multi-view images. However, the lack of 3D awareness often leads to inconsistent results, necessitating costly per-scene optimization to enforce 3D consistency. In this work, we introduce MV-SAM, a framework for multi-view segmentation that achieves 3D consistency using pointmaps -- 3D points reconstructed from unposed images by recent visual geometry models. Leveraging the pixel-point one-to-one correspondence of pointmaps, MV-SAM lifts images and prompts into 3D space, eliminating the need for explicit 3D networks or annotated 3D data. Specifically, MV-SAM extends SAM by lifting image embeddings from its pretrained encoder into 3D point embeddings, which are decoded by a transformer using cross-attention with 3D prompt embeddings. This design aligns 2D interactions with 3D geometry, enabling the model to implicitly learn consistent masks across views through 3D positional embeddings. Trained on the SA-1B dataset, our method generalizes well across domains, outperforming SAM2-Video and achieving comparable performance with per-scene optimization baselines on NVOS, SPIn-NeRF, ScanNet++, uCo3D, and DL3DV benchmarks. Code will be released.

Cheng Sun, Zhen Zhang, Shaofu Yang

In this paper, we study a second-order approach to policy optimization in reinforcement learning. Existing second-order methods often suffer from suboptimal sample complexity or rely on unrealistic assumptions about importance sampling. To overcome these limitations, we propose VR-CR-PN, a variance-reduced cubic-regularized policy Newton algorithm. To the best of our knowledge, this is the first algorithm that integrates Hessian-aided variance reduction with second-order policy optimization, effectively addressing the distribution shift problem and achieving best-known sample complexity under general nonconvex conditions but without the need for importance sampling. We theoretically establish that VR-CR-PN achieves a sample complexity of $\tilde{\mathcal{O}}(ε^{-3})$ to reach an $ε$-second-order stationary point, significantly improving upon the previous best result of $\tilde{\mathcal{O}}(ε^{-3.5})$ under comparable assumptions. As an additional contribution, we introduce a novel Hessian estimator for the expected return function, which admits a uniform upper bound independent of the horizon length $H$, allowing the algorithm to achieve horizon-independent sample complexity.

Wei-En Tai, Yu-Lin Shih, Cheng Sun et al. · nvidia

Amodal instance segmentation, which aims to detect and segment both visible and invisible parts of objects in images, plays a crucial role in various applications including autonomous driving, robotic manipulation, and scene understanding. While existing methods require training both front-end detectors and mask decoders jointly, this approach lacks flexibility and fails to leverage the strengths of pre-existing modal detectors. To address this limitation, we propose SAMEO, a novel framework that adapts the Segment Anything Model (SAM) as a versatile mask decoder capable of interfacing with various front-end detectors to enable mask prediction even for partially occluded objects. Acknowledging the constraints of limited amodal segmentation datasets, we introduce Amodal-LVIS, a large-scale synthetic dataset comprising 300K images derived from the modal LVIS and LVVIS datasets. This dataset significantly expands the training data available for amodal segmentation research. Our experimental results demonstrate that our approach, when trained on the newly extended dataset, including Amodal-LVIS, achieves remarkable zero-shot performance on both COCOA-cls and D2SA benchmarks, highlighting its potential for generalization to unseen scenarios.

Ting-Hsun Chi, Chu-Rong Chen, Chi-Tun Hsu et al.

Reconstructing accurate surfaces with radiance fields has progressed rapidly, yet two promising explicit representations, 3D Gaussian Splatting and sparse-voxel rasterization, exhibit complementary strengths and weaknesses. 3D Gaussian Splatting converges quickly and carries useful geometric priors, but surface fidelity is limited by its point-like parameterization. Sparse-voxel rasterization provides continuous opacity fields and crisp geometry, but its typical uniform dense-grid initialization slows convergence and underutilizes scene structure. We combine the advantages of both by introducing a voxel initialization method that places voxels at plausible locations and with appropriate levels of detail, yielding a strong starting point for per-scene optimization. To further enhance depth consistency without blurring edges, we propose refined depth geometry supervision that converts multi-view cues into direct per-ray depth regularization. Experiments on standard benchmarks demonstrate improvements over prior methods in geometric accuracy, better fine-structure recovery, and more complete surfaces, while maintaining fast convergence.

Yu-Lin Shih, Wei-En Tai, Cheng Sun et al.

We introduce a new task, Referring and Reasoning for Selective Masks (R2SM), which extends text-guided segmentation by incorporating mask-type selection driven by user intent. This task challenges vision-language models to determine whether to generate a modal (visible) or amodal (complete) segmentation mask based solely on natural language prompts. To support the R2SM task, we present the R2SM dataset, constructed by augmenting annotations of COCOA-cls, D2SA, and MUVA. The R2SM dataset consists of both modal and amodal text queries, each paired with the corresponding ground-truth mask, enabling model finetuning and evaluation for the ability to segment images as per user intent. Specifically, the task requires the model to interpret whether a given prompt refers to only the visible part of an object or to its complete shape, including occluded regions, and then produce the appropriate segmentation. For example, if a prompt explicitly requests the whole shape of a partially hidden object, the model is expected to output an amodal mask that completes the occluded parts. In contrast, prompts without explicit mention of hidden regions should generate standard modal masks. The R2SM benchmark provides a challenging and insightful testbed for advancing research in multimodal reasoning and intent-aware segmentation.

Ding-Jiun Huang, Zi-Ting Chou, Yu-Chiang Frank Wang et al.

NeRF-based methods reconstruct 3D scenes by building a radiance field with implicit or explicit representations. While NeRF-based methods can perform novel view synthesis (NVS) at arbitrary scale, the performance in high-resolution novel view synthesis (HRNVS) with low-resolution (LR) optimization often results in oversmoothing. On the other hand, single-image super-resolution (SR) aims to enhance LR images to HR counterparts but lacks multi-view consistency. To address these challenges, we propose Arbitrary-Scale Super-Resolution NeRF (ASSR-NeRF), a novel framework for super-resolution novel view synthesis (SRNVS). We propose an attention-based VoxelGridSR model to directly perform 3D super-resolution (SR) on the optimized volume. Our model is trained on diverse scenes to ensure generalizability. For unseen scenes trained with LR views, we then can directly apply our VoxelGridSR to further refine the volume and achieve multi-view consistent SR. We demonstrate quantitative and qualitatively that the proposed method achieves significant performance in SRNVS.

Linqing Chen, Weilei Wang, Zilong Bai et al.

Large language models (LLMs) have revolutionized Natural Language Processing (NLP) by minimizing the need for complex feature engineering. However, the application of LLMs in specialized domains like biopharmaceuticals and chemistry remains largely unexplored. These fields are characterized by intricate terminologies, specialized knowledge, and a high demand for precision areas where general purpose LLMs often fall short. In this study, we introduce PharmaGPT, a suite of domain specilized LLMs with 13 billion and 70 billion parameters, specifically trained on a comprehensive corpus tailored to the Bio-Pharmaceutical and Chemical domains. Our evaluation shows that PharmaGPT surpasses existing general models on specific-domain benchmarks such as NAPLEX, demonstrating its exceptional capability in domain-specific tasks. Remarkably, this performance is achieved with a model that has only a fraction, sometimes just one-tenth-of the parameters of general-purpose large models. This advancement establishes a new benchmark for LLMs in the bio-pharmaceutical and chemical fields, addressing the existing gap in specialized language modeling. It also suggests a promising path for enhanced research and development, paving the way for more precise and effective NLP applications in these areas.

Cheng Sun, Min Sun, Hwann-Tzong Chen

We present a super-fast convergence approach to reconstructing the per-scene radiance field from a set of images that capture the scene with known poses. This task, which is often applied to novel view synthesis, is recently revolutionized by Neural Radiance Field (NeRF) for its state-of-the-art quality and flexibility. However, NeRF and its variants require a lengthy training time ranging from hours to days for a single scene. In contrast, our approach achieves NeRF-comparable quality and converges rapidly from scratch in less than 15 minutes with a single GPU. We adopt a representation consisting of a density voxel grid for scene geometry and a feature voxel grid with a shallow network for complex view-dependent appearance. Modeling with explicit and discretized volume representations is not new, but we propose two simple yet non-trivial techniques that contribute to fast convergence speed and high-quality output. First, we introduce the post-activation interpolation on voxel density, which is capable of producing sharp surfaces in lower grid resolution. Second, direct voxel density optimization is prone to suboptimal geometry solutions, so we robustify the optimization process by imposing several priors. Finally, evaluation on five inward-facing benchmarks shows that our method matches, if not surpasses, NeRF's quality, yet it only takes about 15 minutes to train from scratch for a new scene.

Chi-Wei Hsiao, Cheng Sun, Hwann-Tzong Chen et al.

We present a novel pyramidal output representation to ensure parsimony with our "specialize and fuse" process for semantic segmentation. A pyramidal "output" representation consists of coarse-to-fine levels, where each level is "specialize" in a different class distribution (e.g., more stuff than things classes at coarser levels). Two types of pyramidal outputs (i.e., unity and semantic pyramid) are "fused" into the final semantic output, where the unity pyramid indicates unity-cells (i.e., all pixels in such cell share the same semantic label). The process ensures parsimony by predicting a relatively small number of labels for unity-cells (e.g., a large cell of grass) to build the final semantic output. In addition to the "output" representation, we design a coarse-to-fine contextual module to aggregate the "features" representation from different levels. We validate the effectiveness of each key module in our method through comprehensive ablation studies. Finally, our approach achieves state-of-the-art performance on three widely-used semantic segmentation datasets -- ADE20K, COCO-Stuff, and Pascal-Context.

Cheng Sun, Chi-Wei Hsiao, Ning-Hsu Wang et al.

Indoor panorama typically consists of human-made structures parallel or perpendicular to gravity. We leverage this phenomenon to approximate the scene in a 360-degree image with (H)orizontal-planes and (V)ertical-planes. To this end, we propose an effective divide-and-conquer strategy that divides pixels based on their plane orientation estimation; then, the succeeding instance segmentation module conquers the task of planes clustering more easily in each plane orientation group. Besides, parameters of V-planes depend on camera yaw rotation, but translation-invariant CNNs are less aware of the yaw change. We thus propose a yaw-invariant V-planar reparameterization for CNNs to learn. We create a benchmark for indoor panorama planar reconstruction by extending existing 360 depth datasets with ground truth H\&V-planes (referred to as PanoH&V dataset) and adopt state-of-the-art planar reconstruction methods to predict H\&V-planes as our baselines. Our method outperforms the baselines by a large margin on the proposed dataset.

Ching-Yu Hsu, Cheng Sun, Hwann-Tzong Chen

We present Omnidirectional Neural Radiance Fields (OmniNeRF), the first method to the application of parallax-enabled novel panoramic view synthesis. Recent works for novel view synthesis focus on perspective images with limited field-of-view and require sufficient pictures captured in a specific condition. Conversely, OmniNeRF can generate panorama images for unknown viewpoints given a single equirectangular image as training data. To this end, we propose to augment the single RGB-D panorama by projecting back and forth between a 3D world and different 2D panoramic coordinates at different virtual camera positions. By doing so, we are able to optimize an Omnidirectional Neural Radiance Field with visible pixels collecting from omnidirectional viewing angles at a fixed center for the estimation of new viewing angles from varying camera positions. As a result, the proposed OmniNeRF achieves convincing renderings of novel panoramic views that exhibit the parallax effect. We showcase the effectiveness of each of our proposals on both synthetic and real-world datasets.

Yicheng Deng, Cheng Sun, Jiahui Zhu et al.

Recovering 3D human pose from 2D joints is still a challenging problem, especially without any 3D annotation, video information, or multi-view information. In this paper, we present an unsupervised GAN-based model consisting of multiple weight-sharing generators to estimate a 3D human pose from a single image without 3D annotations. In our model, we introduce single-view-multi-angle consistency (SVMAC) to significantly improve the estimation performance. With 2D joint locations as input, our model estimates a 3D pose and a camera simultaneously. During training, the estimated 3D pose is rotated by random angles and the estimated camera projects the rotated 3D poses back to 2D. The 2D reprojections will be fed into weight-sharing generators to estimate the corresponding 3D poses and cameras, which are then mixed to impose SVMAC constraints to self-supervise the training process. The experimental results show that our method outperforms the state-of-the-art unsupervised methods on Human 3.6M and MPI-INF-3DHP. Moreover, qualitative results on MPII and LSP show that our method can generalize well to unknown data.

Yicheng Deng, Cheng Sun, Yongqi Sun et al.

3D human pose estimation from a single image is still a challenging problem despite the large amount of work that has been performed in this field. Generally, most methods directly use neural networks and ignore certain constraints (e.g., reprojection constraints, joint angle, and bone length constraints). While a few methods consider these constraints but train the network separately, they cannot effectively solve the depth ambiguity problem. In this paper, we propose a GAN-based model for 3D human pose estimation, in which a reprojection network is employed to learn the mapping of the distribution from 3D poses to 2D poses, and a discriminator is employed for 2D-3D consistency discrimination. We adopt a novel strategy to synchronously train the generator, the reprojection network and the discriminator. Furthermore, inspired by the typical kinematic chain space (KCS) matrix, we introduce a weighted KCS matrix and take it as one of the discriminator's inputs to impose joint angle and bone length constraints. The experimental results on Human3.6M show that our method significantly outperforms state-of-the-art methods in most cases.

Cheng Sun, Min Sun, Hwann-Tzong Chen

We present HoHoNet, a versatile and efficient framework for holistic understanding of an indoor 360-degree panorama using a Latent Horizontal Feature (LHFeat). The compact LHFeat flattens the features along the vertical direction and has shown success in modeling per-column modality for room layout reconstruction. HoHoNet advances in two important aspects. First, the deep architecture is redesigned to run faster with improved accuracy. Second, we propose a novel horizon-to-dense module, which relaxes the per-column output shape constraint, allowing per-pixel dense prediction from LHFeat. HoHoNet is fast: It runs at 52 FPS and 110 FPS with ResNet-50 and ResNet-34 backbones respectively, for modeling dense modalities from a high-resolution $512 \times 1024$ panorama. HoHoNet is also accurate. On the tasks of layout estimation and semantic segmentation, HoHoNet achieves results on par with current state-of-the-art. On dense depth estimation, HoHoNet outperforms all the prior arts by a large margin.

Shih-Han Chou, Cheng Sun, Wen-Yen Chang et al.

While there are several widely used object detection datasets, current computer vision algorithms are still limited in conventional images. Such images narrow our vision in a restricted region. On the other hand, 360° images provide a thorough sight. In this paper, our goal is to provide a standard dataset to facilitate the vision and machine learning communities in 360° domain. To facilitate the research, we present a real-world 360° panoramic object detection dataset, 360-Indoor, which is a new benchmark for visual object detection and class recognition in 360° indoor images. It is achieved by gathering images of complex indoor scenes containing common objects and the intensive annotated bounding field-of-view. In addition, 360-Indoor has several distinct properties: (1) the largest category number (37 labels in total). (2) the most complete annotations on average (27 bounding boxes per image). The selected 37 objects are all common in indoor scene. With around 3k images and 90k labels in total, 360-Indoor achieves the largest dataset for detection in 360° images. In the end, extensive experiments on the state-of-the-art methods for both classification and detection are provided. We will release this dataset in the near future.

Chi-Wei Hsiao, Cheng Sun, Min Sun et al.

This paper proposes a new approach, Flat2Layout, for estimating general indoor room layout from a single-view RGB image whereas existing methods can only produce layout topologies captured from the box-shaped room. The proposed flat representation encodes the layout information into row vectors which are treated as the training target of the deep model. A dynamic programming based postprocessing is employed to decode the estimated flat output from the deep model into the final room layout. Flat2Layout achieves state-of-the-art performance on existing room layout benchmark. This paper also constructs a benchmark for validating the performance on general layout topologies, where Flat2Layout achieves good performance on general room types. Flat2Layout is applicable on more scenario for layout estimation and would have an impact on applications of Scene Modeling, Robotics, and Augmented Reality.

Cheng Sun, Chi-Wei Hsiao, Min Sun et al.

We present a new approach to the problem of estimating the 3D room layout from a single panoramic image. We represent room layout as three 1D vectors that encode, at each image column, the boundary positions of floor-wall and ceiling-wall, and the existence of wall-wall boundary. The proposed network, HorizonNet, trained for predicting 1D layout, outperforms previous state-of-the-art approaches. The designed post-processing procedure for recovering 3D room layouts from 1D predictions can automatically infer the room shape with low computation cost - it takes less than 20ms for a panorama image while prior works might need dozens of seconds. We also propose Pano Stretch Data Augmentation, which can diversify panorama data and be applied to other panorama-related learning tasks. Due to the limited data available for non-cuboid layout, we relabel 65 general layout from the current dataset for finetuning. Our approach shows good performance on general layouts by qualitative results and cross-validation.

Jie Liu, Cheng Sun, Xiang Xu et al.

The video-based person re-identification is to recognize a person under different cameras, which is a crucial task applied in visual surveillance system. Most previous methods mainly focused on the feature of full body in the frame. In this paper we propose a novel Spatial and Temporal Features Mixture Model (STFMM) based on convolutional neural network (CNN) and recurrent neural network (RNN), in which the human body is split into $N$ parts in horizontal direction so that we can obtain more specific features. The proposed method skillfully integrates features of each part to achieve more expressive representation of each person. We first split the video sequence into $N$ part sequences which include the information of head, waist, legs and so on. Then the features are extracted by STFMM whose $2N$ inputs are obtained from the developed Siamese network, and these features are combined into a discriminative representation for one person. Experiments are conducted on the iLIDS-VID and PRID-2011 datasets. The results demonstrate that our approach outperforms existing methods for video-based person re-identification. It achieves a rank-1 CMC accuracy of 74\% on the iLIDS-VID dataset, exceeding the the most recently developed method ASTPN by 12\%. For the cross-data testing, our method achieves a rank-1 CMC accuracy of 48\% exceeding the ASTPN method by 18\%, which shows that our model has significant stability.