Peter Kontschieder

Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang et al. · mit

Humans can orient themselves in their 3D environments using simple 2D maps. Differently, algorithms for visual localization mostly rely on complex 3D point clouds that are expensive to build, store, and maintain over time. We bridge this gap by introducing OrienterNet, the first deep neural network that can localize an image with sub-meter accuracy using the same 2D semantic maps that humans use. OrienterNet estimates the location and orientation of a query image by matching a neural Bird's-Eye View with open and globally available maps from OpenStreetMap, enabling anyone to localize anywhere such maps are available. OrienterNet is supervised only by camera poses but learns to perform semantic matching with a wide range of map elements in an end-to-end manner. To enable this, we introduce a large crowd-sourced dataset of images captured across 12 cities from the diverse viewpoints of cars, bikes, and pedestrians. OrienterNet generalizes to new datasets and pushes the state of the art in both robotics and AR scenarios. The code and trained model will be released publicly.

Norman Müller, Yawar Siddiqui, Lorenzo Porzi et al.

We introduce DiffRF, a novel approach for 3D radiance field synthesis based on denoising diffusion probabilistic models. While existing diffusion-based methods operate on images, latent codes, or point cloud data, we are the first to directly generate volumetric radiance fields. To this end, we propose a 3D denoising model which directly operates on an explicit voxel grid representation. However, as radiance fields generated from a set of posed images can be ambiguous and contain artifacts, obtaining ground truth radiance field samples is non-trivial. We address this challenge by pairing the denoising formulation with a rendering loss, enabling our model to learn a deviated prior that favours good image quality instead of trying to replicate fitting errors like floating artifacts. In contrast to 2D-diffusion models, our model learns multi-view consistent priors, enabling free-view synthesis and accurate shape generation. Compared to 3D GANs, our diffusion-based approach naturally enables conditional generation such as masked completion or single-view 3D synthesis at inference time.

Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Buló et al.

We propose Panoptic Lifting, a novel approach for learning panoptic 3D volumetric representations from images of in-the-wild scenes. Once trained, our model can render color images together with 3D-consistent panoptic segmentation from novel viewpoints. Unlike existing approaches which use 3D input directly or indirectly, our method requires only machine-generated 2D panoptic segmentation masks inferred from a pre-trained network. Our core contribution is a panoptic lifting scheme based on a neural field representation that generates a unified and multi-view consistent, 3D panoptic representation of the scene. To account for inconsistencies of 2D instance identifiers across views, we solve a linear assignment with a cost based on the model's current predictions and the machine-generated segmentation masks, thus enabling us to lift 2D instances to 3D in a consistent way. We further propose and ablate contributions that make our method more robust to noisy, machine-generated labels, including test-time augmentations for confidence estimates, segment consistency loss, bounded segmentation fields, and gradient stopping. Experimental results validate our approach on the challenging Hypersim, Replica, and ScanNet datasets, improving by 8.4, 13.8, and 10.6% in scene-level PQ over state of the art.

Norman Müller, Andrea Simonelli, Lorenzo Porzi et al.

We introduce AutoRF - a new approach for learning neural 3D object representations where each object in the training set is observed by only a single view. This setting is in stark contrast to the majority of existing works that leverage multiple views of the same object, employ explicit priors during training, or require pixel-perfect annotations. To address this challenging setting, we propose to learn a normalized, object-centric representation whose embedding describes and disentangles shape, appearance, and pose. Each encoding provides well-generalizable, compact information about the object of interest, which is decoded in a single-shot into a new target view, thus enabling novel view synthesis. We further improve the reconstruction quality by optimizing shape and appearance codes at test time by fitting the representation tightly to the input image. In a series of experiments, we show that our method generalizes well to unseen objects, even across different datasets of challenging real-world street scenes such as nuScenes, KITTI, and Mapillary Metropolis.

Barbara Roessle, Norman Müller, Lorenzo Porzi et al.

Neural Radiance Fields (NeRF) have shown impressive novel view synthesis results; nonetheless, even thorough recordings yield imperfections in reconstructions, for instance due to poorly observed areas or minor lighting changes. Our goal is to mitigate these imperfections from various sources with a joint solution: we take advantage of the ability of generative adversarial networks (GANs) to produce realistic images and use them to enhance realism in 3D scene reconstruction with NeRFs. To this end, we learn the patch distribution of a scene using an adversarial discriminator, which provides feedback to the radiance field reconstruction, thus improving realism in a 3D-consistent fashion. Thereby, rendering artifacts are repaired directly in the underlying 3D representation by imposing multi-view path rendering constraints. In addition, we condition a generator with multi-resolution NeRF renderings which is adversarially trained to further improve rendering quality. We demonstrate that our approach significantly improves rendering quality, e.g., nearly halving LPIPS scores compared to Nerfacto while at the same time improving PSNR by 1.4dB on the advanced indoor scenes of Tanks and Temples.

Linning Xu, Vasu Agrawal, William Laney et al.

We present an end-to-end system for the high-fidelity capture, model reconstruction, and real-time rendering of walkable spaces in virtual reality using neural radiance fields. To this end, we designed and built a custom multi-camera rig to densely capture walkable spaces in high fidelity and with multi-view high dynamic range images in unprecedented quality and density. We extend instant neural graphics primitives with a novel perceptual color space for learning accurate HDR appearance, and an efficient mip-mapping mechanism for level-of-detail rendering with anti-aliasing, while carefully optimizing the trade-off between quality and speed. Our multi-GPU renderer enables high-fidelity volume rendering of our neural radiance field model at the full VR resolution of dual 2K$\times$2K at 36 Hz on our custom demo machine. We demonstrate the quality of our results on our challenging high-fidelity datasets, and compare our method and datasets to existing baselines. We release our dataset on our project website.

Stefano Esposito, Anpei Chen, Christian Reiser et al.

High-quality view synthesis relies on volume rendering, splatting, or surface rendering. While surface rendering is typically the fastest, it struggles to accurately model fuzzy geometry like hair. In turn, alpha-blending techniques excel at representing fuzzy materials but require an unbounded number of samples per ray (P1). Further overheads are induced by empty space skipping in volume rendering (P2) and sorting input primitives in splatting (P3). We present a novel representation for real-time view synthesis where the (P1) number of sampling locations is small and bounded, (P2) sampling locations are efficiently found via rasterization, and (P3) rendering is sorting-free. We achieve this by representing objects as semi-transparent multi-layer meshes rendered in a fixed order. First, we model surface layers as signed distance function (SDF) shells with optimal spacing learned during training. Then, we bake them as meshes and fit UV textures. Unlike single-surface methods, our multi-layer representation effectively models fuzzy objects. In contrast to volume and splatting-based methods, our approach enables real-time rendering on low-power laptops and smartphones.

Narek Tumanyan, Samuel Rota Bulò, Denis Rozumny et al.

Dynamic scene reconstruction from casual videos has seen recent remarkable progress. Numerous approaches have attempted to overcome the ill-posedness of the task by distilling priors from 2D foundational models and by imposing hand-crafted regularization on the optimized motion. However, these methods struggle to reconstruct scenes from extreme novel viewpoints, especially when highly articulated motions are present. In this paper, we present DRoPS, a novel approach that leverages a static pre-scan of the dynamic object as an explicit geometric and appearance prior. While existing state-of-the-art methods fail to fully exploit the pre-scan, DRoPS leverages our novel setup to effectively constrain the solution space and ensure geometrical consistency throughout the sequence. The core of our novelty is twofold: first, we establish a grid-structured and surface-aligned model by organizing Gaussian primitives into pixel grids anchored to the object surface. Second, by leveraging the grid structure of our primitives, we parameterize motion using a CNN conditioned on those grids, injecting strong implicit regularization and correlating the motion of nearby points. Extensive experiments demonstrate that our method significantly outperforms the current state of the art in rendering quality and 3D tracking accuracy.

Ruofan Liang, Norman Müller, Ethan Weber et al.

We present a novel approach for interactive light editing in indoor scenes from a single multi-view scene capture. Our method leverages a generative image-based light decomposition model that factorizes complex indoor scene illumination into its constituent light sources. This factorization enables independent manipulation of individual light sources, specifically allowing control over their state (on/off), chromaticity, and intensity. We further introduce multi-view lighting harmonization to ensure consistent propagation of the lighting decomposition across all scene views. This is integrated into a relightable 3D Gaussian splatting representation, providing real-time interactive control over the individual light sources. Our results demonstrate highly photorealistic lighting decomposition and relighting outcomes across diverse indoor scenes. We evaluate our method on both synthetic and real-world datasets and provide a quantitative and qualitative comparison to state-of-the-art techniques. For video results and interactive demos, see https://luxremix.github.io.

Naama Pearl, Stefano Esposito, Haofei Xu et al.

3D Gaussian Splatting (3DGS) optimization is most commonly performed using standard optimizers (Adam, SGD). While stable across diverse scenes, standard optimizers are general-purpose and not tailored to the structure of the problem. In particular, they produce independent parameter updates that do not capture the structural and spatial relationships within a scene, leading to inefficient optimization and slow convergence. Recent works introduced learned optimizers that predict correlated updates informed by inter-parameter and inter-Gaussian dependencies. However, these methods are trained for a fixed number of optimization iterations and rely on manually scheduled learning rates to avoid degradation. In this paper, we introduce a learned optimizer for 3DGS that avoids degradation over extended optimization horizons without auxiliary mechanisms. To enable this, we propose a meta-learning scheme that extends the optimization horizon via a checkpoint buffer and an optimizer rollout strategy, combined with an architecture that encodes gradient scale information in its latent states. Results show improved early novel view synthesis quality while remaining stable over long horizons, with zero-shot generalization to unseen reconstruction settings. To support our findings, we introduce the first unified framework for training and evaluating both learned and conventional optimizers across sparse and dense view settings. Code and models will be released publicly. Our project page is available at https://naamapearl.github.io/learn2splat .

Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder

In this work we present In-Place Activated Batch Normalization (InPlace-ABN) - a novel approach to drastically reduce the training memory footprint of modern deep neural networks in a computationally efficient way. Our solution substitutes the conventionally used succession of BatchNorm + Activation layers with a single plugin layer, hence avoiding invasive framework surgery while providing straightforward applicability for existing deep learning frameworks. We obtain memory savings of up to 50% by dropping intermediate results and by recovering required information during the backward pass through the inversion of stored forward results, with only minor increase (0.8-2%) in computation time. Also, we demonstrate how frequently used checkpointing approaches can be made computationally as efficient as InPlace-ABN. In our experiments on image classification, we demonstrate on-par results on ImageNet-1k with state-of-the-art approaches. On the memory-demanding task of semantic segmentation, we report results for COCO-Stuff, Cityscapes and Mapillary Vistas, obtaining new state-of-the-art results on the latter without additional training data but in a single-scale and -model scenario. Code can be found at https://github.com/mapillary/inplace_abn .

Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder

In this paper, we address the limitations of Adaptive Density Control (ADC) in 3D Gaussian Splatting (3DGS), a scene representation method achieving high-quality, photorealistic results for novel view synthesis. ADC has been introduced for automatic 3D point primitive management, controlling densification and pruning, however, with certain limitations in the densification logic. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS, leveraging an auxiliary, per-pixel error function as the criterion for densification. We further introduce a mechanism to control the total number of primitives generated per scene and correct a bias in the current opacity handling strategy of ADC during cloning operations. Our approach leads to consistent quality improvements across a variety of benchmark scenes, without sacrificing the method's efficiency.

Haithem Turki, Vasu Agrawal, Samuel Rota Bulò et al.

Neural radiance fields provide state-of-the-art view synthesis quality but tend to be slow to render. One reason is that they make use of volume rendering, thus requiring many samples (and model queries) per ray at render time. Although this representation is flexible and easy to optimize, most real-world objects can be modeled more efficiently with surfaces instead of volumes, requiring far fewer samples per ray. This observation has spurred considerable progress in surface representations such as signed distance functions, but these may struggle to model semi-opaque and thin structures. We propose a method, HybridNeRF, that leverages the strengths of both representations by rendering most objects as surfaces while modeling the (typically) small fraction of challenging regions volumetrically. We evaluate HybridNeRF against the challenging Eyeful Tower dataset along with other commonly used view synthesis datasets. When comparing to state-of-the-art baselines, including recent rasterization-based approaches, we improve error rates by 15-30% while achieving real-time framerates (at least 36 FPS) for virtual-reality resolutions (2Kx2K).

Barbara Roessle, Norman Müller, Lorenzo Porzi et al.

We propose L3DG, the first approach for generative 3D modeling of 3D Gaussians through a latent 3D Gaussian diffusion formulation. This enables effective generative 3D modeling, scaling to generation of entire room-scale scenes which can be very efficiently rendered. To enable effective synthesis of 3D Gaussians, we propose a latent diffusion formulation, operating in a compressed latent space of 3D Gaussians. This compressed latent space is learned by a vector-quantized variational autoencoder (VQ-VAE), for which we employ a sparse convolutional architecture to efficiently operate on room-scale scenes. This way, the complexity of the costly generation process via diffusion is substantially reduced, allowing higher detail on object-level generation, as well as scalability to large scenes. By leveraging the 3D Gaussian representation, the generated scenes can be rendered from arbitrary viewpoints in real-time. We demonstrate that our approach significantly improves visual quality over prior work on unconditional object-level radiance field synthesis and showcase its applicability to room-scale scene generation.

François Darmon, Lorenzo Porzi, Samuel Rota-Bulò et al.

In this paper, we address common error sources for 3D Gaussian Splatting (3DGS) including blur, imperfect camera poses, and color inconsistencies, with the goal of improving its robustness for practical applications like reconstructions from handheld phone captures. Our main contribution involves modeling motion blur as a Gaussian distribution over camera poses, allowing us to address both camera pose refinement and motion blur correction in a unified way. Additionally, we propose mechanisms for defocus blur compensation and for addressing color in-consistencies caused by ambient light, shadows, or due to camera-related factors like varying white balancing settings. Our proposed solutions integrate in a seamless way with the 3DGS formulation while maintaining its benefits in terms of training efficiency and rendering speed. We experimentally validate our contributions on relevant benchmark datasets including Scannet++ and Deblur-NeRF, obtaining state-of-the-art results and thus consistent improvements over relevant baselines.

Tobias Fischer, Lorenzo Porzi, Samuel Rota Bulò et al.

We estimate the radiance field of large-scale dynamic areas from multiple vehicle captures under varying environmental conditions. Previous works in this domain are either restricted to static environments, do not scale to more than a single short video, or struggle to separately represent dynamic object instances. To this end, we present a novel, decomposable radiance field approach for dynamic urban environments. We propose a multi-level neural scene graph representation that scales to thousands of images from dozens of sequences with hundreds of fast-moving objects. To enable efficient training and rendering of our representation, we develop a fast composite ray sampling and rendering scheme. To test our approach in urban driving scenarios, we introduce a new, novel view synthesis benchmark. We show that our approach outperforms prior art by a significant margin on both established and our proposed benchmark while being faster in training and rendering.

Nikhil Keetha, Norman Müller, Johannes Schönberger et al.

We introduce MapAnything, a unified transformer-based feed-forward model that ingests one or more images along with optional geometric inputs such as camera intrinsics, poses, depth, or partial reconstructions, and then directly regresses the metric 3D scene geometry and cameras. MapAnything leverages a factored representation of multi-view scene geometry, i.e., a collection of depth maps, local ray maps, camera poses, and a metric scale factor that effectively upgrades local reconstructions into a globally consistent metric frame. Standardizing the supervision and training across diverse datasets, along with flexible input augmentation, enables MapAnything to address a broad range of 3D vision tasks in a single feed-forward pass, including uncalibrated structure-from-motion, calibrated multi-view stereo, monocular depth estimation, camera localization, depth completion, and more. We provide extensive experimental analyses and model ablations demonstrating that MapAnything outperforms or matches specialist feed-forward models while offering more efficient joint training behavior, thus paving the way toward a universal 3D reconstruction backbone.

Katja Schwarz, Norman Mueller, Peter Kontschieder

Synthesizing consistent and photorealistic 3D scenes is an open problem in computer vision. Video diffusion models generate impressive videos but cannot directly synthesize 3D representations, i.e., lack 3D consistency in the generated sequences. In addition, directly training generative 3D models is challenging due to a lack of 3D training data at scale. In this work, we present Generative Gaussian Splatting (GGS) -- a novel approach that integrates a 3D representation with a pre-trained latent video diffusion model. Specifically, our model synthesizes a feature field parameterized via 3D Gaussian primitives. The feature field is then either rendered to feature maps and decoded into multi-view images, or directly upsampled into a 3D radiance field. We evaluate our approach on two common benchmark datasets for scene synthesis, RealEstate10K and ScanNet+, and find that our proposed GGS model significantly improves both the 3D consistency of the generated multi-view images, and the quality of the generated 3D scenes over all relevant baselines. Compared to a similar model without 3D representation, GGS improves FID on the generated 3D scenes by ~20% on both RealEstate10K and ScanNet+. Project page: https://katjaschwarz.github.io/ggs/

Tobias Fischer, Samuel Rota Bulò, Yung-Hsu Yang et al.

3D Gaussian splatting enables high-quality novel view synthesis (NVS) at real-time frame rates. However, its quality drops sharply as we depart from the training views. Thus, dense captures are needed to match the high-quality expectations of applications like Virtual Reality (VR). However, such dense captures are very laborious and expensive to obtain. Existing works have explored using 2D generative models to alleviate this requirement by distillation or generating additional training views. These models typically rely on a noise-to-data generative process conditioned only on a handful of reference input views, leading to hallucinations, inconsistent generation results, and subsequent reconstruction artifacts. Instead, we propose a multi-view, flow matching model that learns a flow to directly connect novel view renderings from possibly sparse reconstructions to renderings that we expect from dense reconstructions. This enables augmenting scene captures with consistent, generated views to improve reconstruction quality. Our model is trained on a novel dataset of 3.6M image pairs and can process up to 45 views at 540x960 resolution (91K tokens) on one H100 GPU in a single forward pass. Our pipeline consistently improves NVS in sparse- and dense-view scenarios, leading to higher-quality reconstructions than prior works across multiple, widely-used NVS benchmarks.

Katja Schwarz, Denys Rozumnyi, Samuel Rota Bulò et al.

We introduce a recipe for generating immersive 3D worlds from a single image by framing the task as an in-context learning problem for 2D inpainting models. This approach requires minimal training and uses existing generative models. Our process involves two steps: generating coherent panoramas using a pre-trained diffusion model and lifting these into 3D with a metric depth estimator. We then fill unobserved regions by conditioning the inpainting model on rendered point clouds, requiring minimal fine-tuning. Tested on both synthetic and real images, our method produces high-quality 3D environments suitable for VR display. By explicitly modeling the 3D structure of the generated environment from the start, our approach consistently outperforms state-of-the-art, video synthesis-based methods along multiple quantitative image quality metrics. Project Page: https://katjaschwarz.github.io/worlds/

Samuel Rota Bulò, Nemanja Bartolovic, Lorenzo Porzi et al.

We present a novel, hardware rasterized rendering approach for ray-based 3D Gaussian Splatting (RayGS), obtaining both fast and high-quality results for novel view synthesis. Our work contains a mathematically rigorous and geometrically intuitive derivation about how to efficiently estimate all relevant quantities for rendering RayGS models, structured with respect to standard hardware rasterization shaders. Our solution is the first enabling rendering RayGS models at sufficiently high frame rates to support quality-sensitive applications like Virtual and Mixed Reality. Our second contribution enables alias-free rendering for RayGS, by addressing MIP-related issues arising when rendering diverging scales during training and testing. We demonstrate significant performance gains, across different benchmark scenes, while retaining state-of-the-art appearance quality of RayGS.

Andrea Simonelli, Norman Müller, Peter Kontschieder

The increasing availability of digital 3D environments, whether through image-based 3D reconstruction, generation, or scans obtained by robots, is driving innovation across various applications. These come with a significant demand for 3D interaction, such as 3D Interactive Segmentation, which is useful for tasks like object selection and manipulation. Additionally, there is a persistent need for solutions that are efficient, precise, and performing well across diverse settings, particularly in unseen environments and with unfamiliar objects. In this work, we introduce a 3D interactive segmentation method that consistently surpasses previous state-of-the-art techniques on both in-domain and out-of-domain datasets. Our simple approach integrates a voxel-based sparse encoder with a lightweight transformer-based decoder that implements implicit click fusion, achieving superior performance and maximizing efficiency. Our method demonstrates substantial improvements on benchmark datasets, including ScanNet, ScanNet++, S3DIS, and KITTI-360, and also on unseen geometric distributions such as the ones obtained by Gaussian Splatting. The project web-page is available at https://simonelli-andrea.github.io/easy3d.

Denys Rozumnyi, Jonathon Luiten, Numair Khan et al.

Transforming casually captured, monocular videos into fully immersive dynamic experiences is a highly ill-posed task, and comes with significant challenges, e.g., reconstructing unseen regions, and dealing with the ambiguity in monocular depth estimation. In this work we introduce BulletGen, an approach that takes advantage of generative models to correct errors and complete missing information in a Gaussian-based dynamic scene representation. This is done by aligning the output of a diffusion-based video generation model with the 4D reconstruction at a single frozen "bullet-time" step. The generated frames are then used to supervise the optimization of the 4D Gaussian model. Our method seamlessly blends generative content with both static and dynamic scene components, achieving state-of-the-art results on both novel-view synthesis, and 2D/3D tracking tasks.

Norman Müller, Katja Schwarz, Barbara Roessle et al.

We introduce MultiDiff, a novel approach for consistent novel view synthesis of scenes from a single RGB image. The task of synthesizing novel views from a single reference image is highly ill-posed by nature, as there exist multiple, plausible explanations for unobserved areas. To address this issue, we incorporate strong priors in form of monocular depth predictors and video-diffusion models. Monocular depth enables us to condition our model on warped reference images for the target views, increasing geometric stability. The video-diffusion prior provides a strong proxy for 3D scenes, allowing the model to learn continuous and pixel-accurate correspondences across generated images. In contrast to approaches relying on autoregressive image generation that are prone to drifts and error accumulation, MultiDiff jointly synthesizes a sequence of frames yielding high-quality and multi-view consistent results -- even for long-term scene generation with large camera movements, while reducing inference time by an order of magnitude. For additional consistency and image quality improvements, we introduce a novel, structured noise distribution. Our experimental results demonstrate that MultiDiff outperforms state-of-the-art methods on the challenging, real-world datasets RealEstate10K and ScanNet. Finally, our model naturally supports multi-view consistent editing without the need for further tuning.

Jun-Kun Chen, Samuel Rota Bulò, Norman Müller et al.

This paper proposes ConsistDreamer - a novel framework that lifts 2D diffusion models with 3D awareness and 3D consistency, thus enabling high-fidelity instruction-guided scene editing. To overcome the fundamental limitation of missing 3D consistency in 2D diffusion models, our key insight is to introduce three synergetic strategies that augment the input of the 2D diffusion model to become 3D-aware and to explicitly enforce 3D consistency during the training process. Specifically, we design surrounding views as context-rich input for the 2D diffusion model, and generate 3D-consistent, structured noise instead of image-independent noise. Moreover, we introduce self-supervised consistency-enforcing training within the per-scene editing procedure. Extensive evaluation shows that our ConsistDreamer achieves state-of-the-art performance for instruction-guided scene editing across various scenes and editing instructions, particularly in complicated large-scale indoor scenes from ScanNet++, with significantly improved sharpness and fine-grained textures. Notably, ConsistDreamer stands as the first work capable of successfully editing complex (e.g., plaid/checkered) patterns. Our project page is at immortalco.github.io/ConsistDreamer.

Tobias Fischer, Jonas Kulhanek, Samuel Rota Bulò et al.

We present an efficient neural 3D scene representation for novel-view synthesis (NVS) in large-scale, dynamic urban areas. Existing works are not well suited for applications like mixed-reality or closed-loop simulation due to their limited visual quality and non-interactive rendering speeds. Recently, rasterization-based approaches have achieved high-quality NVS at impressive speeds. However, these methods are limited to small-scale, homogeneous data, i.e. they cannot handle severe appearance and geometry variations due to weather, season, and lighting and do not scale to larger, dynamic areas with thousands of images. We propose 4DGF, a neural scene representation that scales to large-scale dynamic urban areas, handles heterogeneous input data, and substantially improves rendering speeds. We use 3D Gaussians as an efficient geometry scaffold while relying on neural fields as a compact and flexible appearance model. We integrate scene dynamics via a scene graph at global scale while modeling articulated motions on a local level via deformations. This decomposed approach enables flexible scene composition suitable for real-world applications. In experiments, we surpass the state-of-the-art by over 3 dB in PSNR and more than 200 times in rendering speed.

Yung-Hsu Yang, Thomas E. Huang, Min Sun et al.

Aggregating information from features across different layers is an essential operation for dense prediction models. Despite its limited expressiveness, feature concatenation dominates the choice of aggregation operations. In this paper, we introduce Attentive Feature Aggregation (AFA) to fuse different network layers with more expressive non-linear operations. AFA exploits both spatial and channel attention to compute weighted average of the layer activations. Inspired by neural volume rendering, we extend AFA with Scale-Space Rendering (SSR) to perform late fusion of multi-scale predictions. AFA is applicable to a wide range of existing network designs. Our experiments show consistent and significant improvements on challenging semantic segmentation benchmarks, including Cityscapes, BDD100K, and Mapillary Vistas, at negligible computational and parameter overhead. In particular, AFA improves the performance of the Deep Layer Aggregation (DLA) model by nearly 6% mIoU on Cityscapes. Our experimental analyses show that AFA learns to progressively refine segmentation maps and to improve boundary details, leading to new state-of-the-art results on boundary detection benchmarks on BSDS500 and NYUDv2. Code and video resources are available at http://vis.xyz/pub/dla-afa.

Idoia Ruiz, Lorenzo Porzi, Samuel Rota Bulò et al.

We introduce the problem of weakly supervised Multi-Object Tracking and Segmentation, i.e. joint weakly supervised instance segmentation and multi-object tracking, in which we do not provide any kind of mask annotation. To address it, we design a novel synergistic training strategy by taking advantage of multi-task learning, i.e. classification and tracking tasks guide the training of the unsupervised instance segmentation. For that purpose, we extract weak foreground localization information, provided by Grad-CAM heatmaps, to generate a partial ground truth to learn from. Additionally, RGB image level information is employed to refine the mask prediction at the edges of the objects. We evaluate our method on KITTI MOTS, the most representative benchmark for this task, reducing the performance gap on the MOTSP metric between the fully supervised and weakly supervised approach to just 12% and 12.7% for cars and pedestrians, respectively.

Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder

Crop-based training strategies decouple training resolution from GPU memory consumption, allowing the use of large-capacity panoptic segmentation networks on multi-megapixel images. Using crops, however, can introduce a bias towards truncating or missing large objects. To address this, we propose a novel crop-aware bounding box regression loss (CABB loss), which promotes predictions to be consistent with the visible parts of the cropped objects, while not over-penalizing them for extending outside of the crop. We further introduce a novel data sampling and augmentation strategy which improves generalization across scales by counteracting the imbalanced distribution of object sizes. Combining these two contributions with a carefully designed, top-down panoptic segmentation architecture, we obtain new state-of-the-art results on the challenging Mapillary Vistas (MVD), Indian Driving and Cityscapes datasets, surpassing the previously best approach on MVD by +4.5% PQ and +5.2% mAP.

Andrea Simonelli, Samuel Rota Bulò, Lorenzo Porzi et al.

Pseudo-LiDAR-based methods for monocular 3D object detection have received considerable attention in the community due to the performance gains exhibited on the KITTI3D benchmark, in particular on the commonly reported validation split. This generated a distorted impression about the superiority of Pseudo-LiDAR-based (PL-based) approaches over methods working with RGB images only. Our first contribution consists in rectifying this view by pointing out and showing experimentally that the validation results published by PL-based methods are substantially biased. The source of the bias resides in an overlap between the KITTI3D object detection validation set and the training/validation sets used to train depth predictors feeding PL-based methods. Surprisingly, the bias remains also after geographically removing the overlap. This leaves the test set as the only reliable set for comparison, where published PL-based methods do not excel. Our second contribution brings PL-based methods back up in the ranking with the design of a novel deep architecture which introduces a 3D confidence prediction module. We show that 3D confidence estimation techniques derived from RGB-only 3D detection approaches can be successfully integrated into our framework and, more importantly, that improved performance can be obtained with a newly designed 3D confidence measure, leading to state-of-the-art performance on the KITTI3D benchmark.

Markus Hofinger, Samuel Rota Bulò, Lorenzo Porzi et al.

In this work we review the coarse-to-fine spatial feature pyramid concept, which is used in state-of-the-art optical flow estimation networks to make exploration of the pixel flow search space computationally tractable and efficient. Within an individual pyramid level, we improve the cost volume construction process by departing from a warping- to a sampling-based strategy, which avoids ghosting and hence enables us to better preserve fine flow details. We further amplify the positive effects through a level-specific, loss max-pooling strategy that adaptively shifts the focus of the learning process on under-performing predictions. Our second contribution revises the gradient flow across pyramid levels. The typical operations performed at each pyramid level can lead to noisy, or even contradicting gradients across levels. We show and discuss how properly blocking some of these gradient components leads to improved convergence and ultimately better performance. Finally, we introduce a distillation concept to counteract the issue of catastrophic forgetting and thus preserving knowledge over models sequentially trained on multiple datasets. Our findings are conceptually simple and easy to implement, yet result in compelling improvements on relevant error measures that we demonstrate via exhaustive ablations on datasets like Flying Chairs2, Flying Things, Sintel and KITTI. We establish new state-of-the-art results on the challenging Sintel and KITTI 2012 test datasets, and even show the portability of our findings to different optical flow and depth from stereo approaches.

Andrea Simonelli, Samuel Rota Bulò, Lorenzo Porzi et al.

While expensive LiDAR and stereo camera rigs have enabled the development of successful 3D object detection methods, monocular RGB-only approaches lag much behind. This work advances the state of the art by introducing MoVi-3D, a novel, single-stage deep architecture for monocular 3D object detection. MoVi-3D builds upon a novel approach which leverages geometrical information to generate, both at training and test time, virtual views where the object appearance is normalized with respect to distance. These virtually generated views facilitate the detection task as they significantly reduce the visual appearance variability associated to objects placed at different distances from the camera. As a consequence, the deep model is relieved from learning depth-specific representations and its complexity can be significantly reduced. In particular, in this work we show that, thanks to our virtual views generation process, a lightweight, single-stage architecture suffices to set new state-of-the-art results on the popular KITTI3D benchmark.

Lorenzo Porzi, Markus Hofinger, Idoia Ruiz et al.

In this work we contribute a novel pipeline to automatically generate training data, and to improve over state-of-the-art multi-object tracking and segmentation (MOTS) methods. Our proposed track mining algorithm turns raw street-level videos into high-fidelity MOTS training data, is scalable and overcomes the need of expensive and time-consuming manual annotation approaches. We leverage state-of-the-art instance segmentation results in combination with optical flow predictions, also trained on automatically harvested training data. Our second major contribution is MOTSNet - a deep learning, tracking-by-detection architecture for MOTS - deploying a novel mask-pooling layer for improved object association over time. Training MOTSNet with our automatically extracted data leads to significantly improved sMOTSA scores on the novel KITTI MOTS dataset (+1.9%/+7.5% on cars/pedestrians), and MOTSNet improves by +4.1% over previously best methods on the MOTSChallenge dataset. Our most impressive finding is that we can improve over previous best-performing works, even in complete absence of manually annotated MOTS training data.

Andrea Simonelli, Samuel Rota Rota Bulò, Lorenzo Porzi et al.

In this paper we propose an approach for monocular 3D object detection from a single RGB image, which leverages a novel disentangling transformation for 2D and 3D detection losses and a novel, self-supervised confidence score for 3D bounding boxes. Our proposed loss disentanglement has the twofold advantage of simplifying the training dynamics in the presence of losses with complex interactions of parameters, and sidestepping the issue of balancing independent regression terms. Our solution overcomes these issues by isolating the contribution made by groups of parameters to a given loss, without changing its nature. We further apply loss disentanglement to another novel, signed Intersection-over-Union criterion-driven loss for improving 2D detection results. Besides our methodological innovations, we critically review the AP metric used in KITTI3D, which emerged as the most important dataset for comparing 3D detection results. We identify and resolve a flaw in the 11-point interpolated AP metric, affecting all previously published detection results and particularly biases the results of monocular 3D detection. We provide extensive experimental evaluations and ablation studies on the KITTI3D and nuScenes datasets, setting new state-of-the-art results on object category car by large margins.

Lorenzo Porzi, Samuel Rota Bulò, Aleksander Colovic et al.

In this work we introduce a novel, CNN-based architecture that can be trained end-to-end to deliver seamless scene segmentation results. Our goal is to predict consistent semantic segmentation and detection results by means of a panoptic output format, going beyond the simple combination of independently trained segmentation and detection models. The proposed architecture takes advantage of a novel segmentation head that seamlessly integrates multi-scale features generated by a Feature Pyramid Network with contextual information conveyed by a light-weight DeepLab-like module. As additional contribution we review the panoptic metric and propose an alternative that overcomes its limitations when evaluating non-instance categories. Our proposed network architecture yields state-of-the-art results on three challenging street-level datasets, i.e. Cityscapes, Indian Driving Dataset and Mapillary Vistas.

Samuel Rota Bulò, Gerhard Neuhold, Peter Kontschieder

We introduce a novel loss max-pooling concept for handling imbalanced training data distributions, applicable as alternative loss layer in the context of deep neural networks for semantic image segmentation. Most real-world semantic segmentation datasets exhibit long tail distributions with few object categories comprising the majority of data and consequently biasing the classifiers towards them. Our method adaptively re-weights the contributions of each pixel based on their observed losses, targeting under-performing classification results as often encountered for under-represented object classes. Our approach goes beyond conventional cost-sensitive learning attempts through adaptive considerations that allow us to indirectly address both, inter- and intra-class imbalances. We provide a theoretical justification of our approach, complementary to experimental analyses on benchmark datasets. In our experiments on the Cityscapes and Pascal VOC 2012 segmentation datasets we find consistently improved results, demonstrating the efficacy of our approach.

Yani Ioannou, Duncan Robertson, Darko Zikic et al.

This paper investigates the connections between two state of the art classifiers: decision forests (DFs, including decision jungles) and convolutional neural networks (CNNs). Decision forests are computationally efficient thanks to their conditional computation property (computation is confined to only a small region of the tree, the nodes along a single branch). CNNs achieve state of the art accuracy, thanks to their representation learning capabilities. We present a systematic analysis of how to fuse conditional computation with representation learning and achieve a continuum of hybrid models with different ratios of accuracy vs. efficiency. We call this new family of hybrid models conditional networks. Conditional networks can be thought of as: i) decision trees augmented with data transformation operators, or ii) CNNs, with block-diagonal sparse weight matrices, and explicit data routing functions. Experimental validation is performed on the common task of image classification on both the CIFAR and Imagenet datasets. Compared to state of the art CNNs, our hybrid models yield the same accuracy with a fraction of the compute cost and much smaller number of parameters.