NeurIPS 2023 Papers

Due to the high price and heavy energy consumption of GPUs, deploying deep models on IoT devices such as microcontrollers makes significant contributions for ecological AI. Conventional methods successfully enable convolutional neural network inference of high resolution images on microcontrollers, while the framework for vision transformers that achieve the state-of-the-art performance in many vision applications still remains unexplored. In this paper, we propose a hardware-algorithm co-optimizations method called MCUFormer to deploy vision transformers on microcontrollers with extremely limited memory, where we jointly design transformer architecture and construct the inference operator library to fit the memory resource constraint. More specifically, we generalize the one-shot network architecture search (NAS) to discover the optimal architecture with highest task performance given the memory budget from the microcontrollers, where we enlarge the existing search space of vision transformers by considering the low-rank decomposition dimensions and patch resolution for memory reduction. For the construction of the inference operator library of vision transformers, we schedule the memory buffer during inference through operator integration, patch embedding decomposition, and token overwriting, allowing the memory buffer to be fully utilized to adapt to the forward pass of the vision transformer. Experimental results demonstrate that our MCUFormer achieves 73.62\% top-1 accuracy on ImageNet for image classification with 320KB memory on STM32F746 microcontroller. Code is available at https://github.com/liangyn22/MCUFormer.

3D object detection in point clouds is important for autonomous driving systems. A primary challenge in 3D object detection stems from the sparse distribution of points within the 3D scene. Existing high-performance methods typically employ 3D sparse convolutional neural networks with small kernels to extract features. To reduce computational costs, these methods resort to submanifold sparse convolutions, which prevent the information exchange among spatially disconnected features. Some recent approaches have attempted to address this problem by introducing large-kernel convolutions or self-attention mechanisms, but they either achieve limited accuracy improvements or incur excessive computational costs. We propose HEDNet, a hierarchical encoder-decoder network for 3D object detection, which leverages encoder-decoder blocks to capture long-range dependencies among features in the spatial space, particularly for large and distant objects. We conducted extensive experiments on the Waymo Open and nuScenes datasets. HEDNet achieved superior detection accuracy on both datasets than previous state-of-the-art methods with competitive efficiency. The code is available at https://github.com/zhanggang001/HEDNet.

Problems involving geometric data arise in physics, chemistry, robotics, computer vision, and many other fields. Such data can take numerous forms, for instance points, direction vectors, translations, or rotations, but to date there is no single architecture that can be applied to such a wide variety of geometric types while respecting their symmetries. In this paper we introduce the Geometric Algebra Transformer (GATr), a general-purpose architecture for geometric data. GATr represents inputs, outputs, and hidden states in the projective geometric (or Clifford) algebra, which offers an efficient 16-dimensional vector-space representation of common geometric objects as well as operators acting on them. GATr is equivariant with respect to E(3), the symmetry group of 3D Euclidean space. As a Transformer, GATr is versatile, efficient, and scalable. We demonstrate GATr in problems from n-body modeling to wall-shear-stress estimation on large arterial meshes to robotic motion planning. GATr consistently outperforms both non-geometric and equivariant baselines in terms of error, data efficiency, and scalability.

3D segmentation of nuclei images is a fundamental task for many biological studies. Despite the rapid advances of large-volume 3D imaging acquisition methods and the emergence of sophisticated algorithms to segment the nuclei in recent years, a benchmark with all cells completely annotated is still missing, making it hard to accurately assess and further improve the performance of the algorithms. The existing nuclei segmentation benchmarks either worked on 2D only or annotated a small number of 3D cells, perhaps due to the high cost of 3D annotation for large-scale data. To fulfill the critical need, we constructed NIS3D, a 3D, high cell density, large-volume, and completely annotated Nuclei Image Segmentation benchmark, assisted by our newly designed semi-automatic annotation software. NIS3D provides more than 22,000 cells across multiple most-used species in this area. Each cell is labeled by three independent annotators, so we can measure the variability of each annotation. A confidence score is computed for each cell, allowing more nuanced testing and performance comparison. A comprehensive review on the methods of segmenting 3D dense nuclei was conducted. The benchmark was used to evaluate the performance of several selected state-of-the-art segmentation algorithms. The best of current methods is still far away from human-level accuracy, corroborating the necessity of generating such a benchmark. The testing results also demonstrated the strength and weakness of each method and pointed out the directions of further methodological development. The dataset can be downloaded here: https://github.com/yu-lab-vt/NIS3D.

Multi-modal Contrastive Representation (MCR) learning aims to encode different modalities into a semantically aligned shared space. This paradigm shows remarkable generalization ability on numerous downstream tasks across various modalities. However, the reliance on massive high-quality data pairs limits its further development on more modalities. This paper proposes a novel training-efficient method for learning MCR without paired data called Connecting Multi-modal Contrastive Representations (C-MCR). Specifically, given two existing MCRs pre-trained on $(\mathcal{A}$, $\mathcal{B})$ and $(\mathcal{B}$, $\mathcal{C})$ modality pairs, we project them to a new space and use the data from the overlapping modality $\mathcal{B}$ to aligning the two MCRs in the new space. Meanwhile, since the modality pairs $(\mathcal{A}$, $\mathcal{B})$ and $(\mathcal{B}$, $\mathcal{C})$ are already aligned within each MCR, the connection learned by overlapping modality can also be transferred to non-overlapping modality pair $(\mathcal{A}$, $\mathcal{C})$. To unleash the potential of C-MCR, we further introduce a semantic-enhanced inter- and intra-MCR connection method. We first enhance the semantic consistency and completion of embeddings across different modalities for more robust alignment. Then we utilize the inter-MCR alignment to establish the connection, and employ the intra-MCR alignment to better maintain the connection for inputs from non-overlapping modalities. To demonstrate the effectiveness of C-MCR, we take the field of audio-visual and 3D-language learning as examples. Specifically, we connect CLIP and CLAP via texts to derive audio-visual representations, and integrate CLIP and ULIP via images for 3D-language representations. Remarkably, without using any paired data, C-MCR for audio-visual achieves state-of-the-art performance on audio-image retrieval, audio-visual source localization, and counterfactual audio-image recognition tasks. Furthermore, C-MCR for 3D-language also attains advanced zero-shot 3D point cloud classification accuracy on ModelNet40. Our project page is available at \url{https://c-mcr.github.io/C-MCR/}

Feedforward generalizable models for implicit shape reconstruction from unoriented point cloud present multiple advantages, including high performance and inference speed. However, they still suffer from generalization issues, ranging from underfitting the input point cloud, to misrepresenting samples outside of the training data distribution, or with toplogies unseen at training. We propose here an efficient mechanism to remedy some of these limitations at test time. We combine the inter-shape data prior of the network with an intra-shape regularization prior of a Nyström Kernel Ridge Regression, that we further adapt by fitting its hyperprameters to the current shape. The resulting shape function defined in a shape specific Reproducing Kernel Hilbert Space benefits from desirable stability and efficiency properties and grants a shape adaptive expressiveness-robustness trade-off. We demonstrate the improvement obtained through our method with respect to baselines and the state-of-the-art using synthetic and real data.

3D object detection (3DOD) from multi-view images is an economically appealing alternative to expensive LiDAR-based detectors, but also an extremely challenging task due to the absence of precise spatial cues. Recent studies have leveraged the teacher-student paradigm for cross-modal distillation, where a strong LiDAR-modality teacher transfers useful knowledge to a multi-view-based image-modality student. However, prior approaches have only focused on minimizing global distances between cross-modal features, which may lead to suboptimal knowledge distillation results. Based on these insights, we propose a novel structural and temporal cross-modal knowledge distillation (STXD) framework for multi-view 3DOD. First, STXD reduces redundancy of the feature components of the student by regularizing the cross-correlation of cross-modal features, while maximizing their similarities. Second, to effectively transfer temporal knowledge, STXD encodes temporal relations of features across a sequence of frames via similarity maps. Lastly, STXD also adopts a response distillation method to further enhance the quality of knowledge distillation at the output-level. Our extensive experiments demonstrate that STXD significantly improves the NDS and mAP of the based student detectors by 2.8%~4.5% on the nuScenes testing dataset.

3D visual grounding, the task of identifying visual objects in 3D scenes based on natural language inputs, plays a critical role in enabling machines to understand and engage with the real-world environment. However, this task is challenging due to the necessity to capture 3D contextual information to distinguish target objects from complex 3D scenes. The absence of annotations for contextual objects and relations further exacerbates the difficulties. In this paper, we propose a novel model, CORE-3DVG, to address these challenges by explicitly learning about contextual objects and relations. Our method accomplishes 3D visual grounding via three sequential modular networks, including a text-guided object detection network, a relation matching network, and a target identification network. During training, we introduce a pseudo-label self-generation strategy and a weakly-supervised method to facilitate the learning of contextual objects and relations, respectively. The proposed techniques allow the networks to focus more effectively on referred objects within 3D scenes by understanding their context better. We validate our model on the challenging Nr3D, Sr3D, and ScanRefer datasets and demonstrate state-of-the-art performance. Our code will be public at https://github.com/yangli18/CORE-3DVG.

Given a set of calibrated images of a scene, we present an approach that produces a simple, compact, and actionable 3D world representation by means of 3D primitives. While many approaches focus on recovering high-fidelity 3D scenes, we focus on parsing a scene into mid-level 3D representations made of a small set of textured primitives. Such representations are interpretable, easy to manipulate and suited for physics-based simulations. Moreover, unlike existing primitive decomposition methods that rely on 3D input data, our approach operates directly on images through differentiable rendering. Specifically, we model primitives as textured superquadric meshes and optimize their parameters from scratch with an image rendering loss. We highlight the importance of modeling transparency for each primitive, which is critical for optimization and also enables handling varying numbers of primitives. We show that the resulting textured primitives faithfully reconstruct the input images and accurately model the visible 3D points, while providing amodal shape completions of unseen object regions. We compare our approach to the state of the art on diverse scenes from DTU, and demonstrate its robustness on real-life captures from BlendedMVS and Nerfstudio. We also showcase how our results can be used to effortlessly edit a scene or perform physical simulations. Code and video results are available at https://www.tmonnier.com/DBW.

Millimetre-wave (mmWave) radar has emerged as an attractive and cost-effective alternative for human activity sensing compared to traditional camera-based systems. mmWave radars are also non-intrusive, providing better protection for user privacy. However, as a Radio Frequency based technology, mmWave radars rely on capturing reflected signals from objects, making them more prone to noise compared to cameras. This raises an intriguing question for the deep learning community: Can we develop more effective point set-based deep learning methods for such attractive sensors? To answer this question, our work, termed MiliPoint, delves into this idea by providing a large-scale, open dataset for the community to explore how mmWave radars can be utilised for human activity recognition. Moreover, MiliPoint stands out as it is larger in size than existing datasets, has more diverse human actions represented, and encompasses all three key tasks in human activity recognition. We have also established a range of point-based deep neural networks such as DGCNN, PointNet++ and PointTransformer, on MiliPoint, which can serve to set the ground baseline for further development.

Given a visual scene, humans have strong intuitions about how a scene can evolve over time under given actions. The intuition, often termed visual intuitive physics, is a critical ability that allows us to make effective plans to manipulate the scene to achieve desired outcomes without relying on extensive trial and error. In this paper, we present a framework capable of learning 3D-grounded visual intuitive physics models from videos of complex scenes with fluids. Our method is composed of a conditional Neural Radiance Field (NeRF)-style visual frontend and a 3D point-based dynamics prediction backend, using which we can impose strong relational and structural inductive bias to capture the structure of the underlying environment. Unlike existing intuitive point-based dynamics works that rely on the supervision of dense point trajectory from simulators, we relax the requirements and only assume access to multi-view RGB images and (imperfect) instance masks acquired using color prior. This enables the proposed model to handle scenarios where accurate point estimation and tracking are hard or impossible. We generate datasets including three challenging scenarios involving fluid, granular materials, and rigid objects in the simulation. The datasets do not include any dense particle information so most previous 3D-based intuitive physics pipelines can barely deal with that. We show our model can make long-horizon future predictions by learning from raw images and significantly outperforms models that do not employ an explicit 3D representation space. We also show that once trained, our model can achieve strong generalization in complex scenarios under extrapolate settings.

Remarkable progress has been made in 3D reconstruction from single-view RGB-D inputs. MCC is the current state-of-the-art method in this field, which achieves unprecedented success by combining vision Transformers with large-scale training. However, we identified two key limitations of MCC: 1) The Transformer decoder is inefficient in handling large number of query points; 2) The 3D representation struggles to recover high-fidelity details. In this paper, we propose a new approach called NU-MCC that addresses these limitations. NU-MCC includes two key innovations: a Neighborhood decoder and a Repulsive Unsigned Distance Function (Repulsive UDF). First, our Neighborhood decoder introduces center points as an efficient proxy of input visual features, allowing each query point to only attend to a small neighborhood. This design not only results in much faster inference speed but also enables the exploitation of finer-scale visual features for improved recovery of 3D textures. Second, our Repulsive UDF is a novel alternative to the occupancy field used in MCC, significantly improving the quality of 3D object reconstruction. Compared to standard UDFs that suffer from holes in results, our proposed Repulsive UDF can achieve more complete surface reconstruction. Experimental results demonstrate that NU-MCC is able to learn a strong 3D representation, significantly advancing the state of the art in single-view 3D reconstruction. Particularly, it outperforms MCC by 9.7% in terms of the F1-score on the CO3D-v2 dataset with more than 5x faster running speed.

A popular benchmark for intra-patient lung registration is provided by the DIR-LAB COPDgene dataset consisting of large-motion in- and expiratory breath-hold CT pairs. This dataset alone, however, does not provide enough samples to properly train state-of-the-art deep learning methods. Other public datasets often also provide only small sample sizes or include primarily small motions between scans that do not translate well to larger deformations. For point-based geometric registration, the PVT1010 dataset provides a large number of vessel point clouds without any correspondences and a labeled test set corresponding to the COPDgene cases. However, the absence of correspondences for supervision complicates training, and a fair comparison with image-based algorithms is infeasible, since CT scans for the training data are not publicly available. We here provide a combined benchmark for image- and point-based registration approaches. We curated a total of 248 public multi-centric in- and expiratory lung CT scans from 124 patients, which show large motion between scans, processed them to ensure sufficient homogeneity between the data and generated vessel point clouds that are well distributed even deeper inside the lungs. For supervised training, we provide vein and artery segmentations of the vessels and multiple thousand image-derived keypoint correspondences for each pair. For validation, we provide multiple scan pairs with manual landmark annotations. Finally, as first baselines on our new benchmark, we evaluate several image and point cloud registration methods on the dataset.

We introduce OpenShape, a method for learning multi-modal joint representations of text, image, and point clouds. We adopt the commonly used multi-modal contrastive learning framework for representation alignment, but with a specific focus on scaling up 3D representations to enable open-world 3D shape understanding. To achieve this, we scale up training data by ensembling multiple 3D datasets and propose several strategies to automatically filter and enrich noisy text descriptions. We also explore and compare strategies for scaling 3D backbone networks and introduce a novel hard negative mining module for more efficient training. We evaluate OpenShape on zero-shot 3D classification benchmarks and demonstrate its superior capabilities for open-world recognition. Specifically, OpenShape achieves a zero-shot accuracy of 46.8% on the 1,156-category Objaverse-LVIS benchmark, compared to less than 10% for existing methods. OpenShape also achieves an accuracy of 85.3% on ModelNet40, outperforming previous zero-shot baseline methods by 20% and performing on par with some fully-supervised methods. Furthermore, we show that our learned embeddings encode a wide range of visual and semantic concepts (e.g., subcategories, color, shape, style) and facilitate fine-grained text-3D and image-3D interactions. Due to their alignment with CLIP embeddings, our learned shape representations can also be integrated with off-the-shelf CLIP-based models for various applications, such as point cloud captioning and point cloud-conditioned image generation.

We propose a novel approach to learning the generative neural fields represented by linear combinations of implicit basis networks. Our algorithm learns basis networks in the form of implicit neural representations and their coefficients in a latent space by either conducting meta-learning or adopting auto-decoding paradigms. The proposed method easily enlarges the capacity of generative neural fields by increasing the number of basis networks while maintaining the size of a network for inference to be small through their weighted model averaging. Consequently, sampling instances using the model is efficient in terms of latency and memory footprint. Moreover, we customize denoising diffusion probabilistic model for a target task to sample latent mixture coefficients, which allows our final model to generate unseen data effectively. Experiments show that our approach achieves competitive generation performance on diverse benchmarks for images, voxel data, and NeRF scenes without sophisticated designs for specific modalities and domains.

We present PrimDiffusion, the first diffusion-based framework for 3D human generation. Devising diffusion models for 3D human generation is difficult due to the intensive computational cost of 3D representations and the articulated topology of 3D humans. To tackle these challenges, our key insight is operating the denoising diffusion process directly on a set of volumetric primitives, which models the human body as a number of small volumes with radiance and kinematic information. This volumetric primitives representation marries the capacity of volumetric representations with the efficiency of primitive-based rendering. Our PrimDiffusion framework has three appealing properties: **1)** compact and expressive parameter space for the diffusion model, **2)** flexible representation that incorporates human prior, and **3)** decoder-free rendering for efficient novel-view and novel-pose synthesis. Extensive experiments validate that PrimDiffusion outperforms state-of-the-art methods in 3D human generation. Notably, compared to GAN-based methods, our PrimDiffusion supports real-time rendering of high-quality 3D humans at a resolution of $512\times512$ once the denoising process is done. We also demonstrate the flexibility of our framework on training-free conditional generation such as texture transfer and 3D inpainting.

Learning accurate and parsimonious point cloud representations of scene surfaces from scratch remains a challenge in 3D representation learning. Existing point-based methods often suffer from the vanishing gradient problem or require a large number of points to accurately model scene geometry and texture. To address these limitations, we propose Proximity Attention Point Rendering (PAPR), a novel method that consists of a point-based scene representation and a differentiable renderer. Our scene representation uses a point cloud where each point is characterized by its spatial position, influence score, and view-independent feature vector. The renderer selects the relevant points for each ray and produces accurate colours using their associated features. PAPR effectively learns point cloud positions to represent the correct scene geometry, even when the initialization drastically differs from the target geometry. Notably, our method captures fine texture details while using only a parsimonious set of points. We also demonstrate four practical applications of our method: zero-shot geometry editing, object manipulation, texture transfer, and exposure control. More results and code are available on our project website at https://zvict.github.io/papr/.

Tone mapping aims to convert high dynamic range (HDR) images to low dynamic range (LDR) representations, a critical task in the camera imaging pipeline. In recent years, 3-Dimensional LookUp Table (3D LUT) based methods have gained attention due to their ability to strike a favorable balance between enhancement performance and computational efficiency. However, these methods often fail to deliver satisfactory results in local areas since the look-up table is a global operator for tone mapping, which works based on pixel values and fails to incorporate crucial local information. To this end, this paper aims to address this issue by exploring a novel strategy that integrates global and local operators by utilizing closed-form Laplacian pyramid decomposition and reconstruction. Specifically, we employ image-adaptive 3D LUTs to manipulate the tone in the low-frequency image by leveraging the specific characteristics of the frequency information. Furthermore, we utilize local Laplacian filters to refine the edge details in the high-frequency components in an adaptive manner. Local Laplacian filters are widely used to preserve edge details in photographs, but their conventional usage involves manual tuning and fixed implementation within camera imaging pipelines or photo editing tools. We propose to learn parameter value maps progressively for local Laplacian filters from annotated data using a lightweight network. Our model achieves simultaneous global tone manipulation and local edge detail preservation in an end-to-end manner. Extensive experimental results on two benchmark datasets demonstrate that the proposed method performs favorably against state-of-the-art methods.

Despite the remarkable achievements of neural radiance fields (NeRF) in representing 3D scenes and generating novel view images, the aliasing issue, rendering 'jaggies' or 'blurry' images at varying camera distances, remains unresolved in most existing approaches. The recently proposed mip-NeRF has effectively addressed this challenge by introducing integrated positional encodings (IPE). However, it relies on MLP architecture to represent the radiance fields, missing out on the fast training speed offered by the latest grid-based methods. In this work, we present mip-Grid, a novel approach that integrates anti-aliasing techniques into grid-based representations for radiance fields, mitigating the aliasing artifacts while enjoying fast training time. Notably, the proposed method uses a single-scale shared grid representation and a single-sampling approach, which only introduces minimal additions to the model parameters and computational costs. To handle scale ambiguity, mip-Grid generates multiple grids by applying simple convolution operations over the shared grid and uses the scale-aware coordinate to retrieve the appropriate features from the generated multiple grids. To test the effectiveness, we incorporated the proposed approach into the two recent representative grid-based methods, TensoRF and K-Planes. The experimental results demonstrated that mip-Grid greatly improved the rendering performance of both methods and showed comparable performance to mip-NeRF on multi-scale datasets while achieving significantly faster training time.