ICLR 2023 Papers

Performant Convolutional Neural Network (CNN) architectures must be tailored to specific tasks in order to consider the length, resolution, and dimensionality of the input data. In this work, we tackle the need for problem-specific CNN architectures. We present the Continuous Convolutional Neural Network (CCNN): a single CNN able to process data of arbitrary resolution, dimensionality and length without any structural changes. Its key component are its continuous convolutional kernels which model long-range dependencies at every layer, and thus remove the need of current CNN architectures for task-dependent downsampling and depths. We showcase the generality of our method by using the same architecture for tasks on sequential ($1{\rm D}$), visual ($2{\rm D}$) and point-cloud ($3{\rm D}$) data. Our CCNN matches and often outperforms the current state-of-the-art across all tasks considered.

Generating photo-realistic video portraits with arbitrary speech audio is a crucial problem in film-making and virtual reality. Recently, several works explore the usage of neural radiance field (NeRF) in this task to improve 3D realness and image fidelity. However, the generalizability of previous NeRF-based methods is limited by the small scale of training data. In this work, we propose GeneFace, a generalized and high-fidelity NeRF-based talking face generation method, which can generate natural results corresponding to various out-of-domain audio. Specifically, we learn a variational motion generator on a large lip-reading corpus, and introduce a domain adaptative post-net to calibrate the result. Moreover, we learn a NeRF-based renderer conditioned on the predicted motion. A head-aware torso-NeRF is proposed to eliminate the head-torso separation problem. Extensive experiments show that our method achieves more generalized and high-fidelity talking face generation compared to previous methods. Video samples and source code are available at https://geneface.github.io .

The recent advance of neural fields, such as neural radiance fields, has significantly pushed the boundary of scene representation learning. Aiming to boost the computation efficiency and rendering quality of 3D scenes, a popular line of research maps the 3D coordinate system to another measuring system, e.g., 2D manifolds and hash tables, for modeling neural fields. The conversion of coordinate systems can be typically dubbed as \emph{gauge transformation}, which is usually a pre-defined mapping function, e.g., orthogonal projection or spatial hash function. This begs a question: can we directly learn a desired gauge transformation along with the neural field in an end-to-end manner? In this work, we extend this problem to a general paradigm with a taxonomy of discrete and continuous cases, and develop an end-to-end learning framework to jointly optimize the gauge transformation and neural fields. To counter the problem that the learning of gauge transformations can collapse easily, we derive a general regularization mechanism from the principle of information conservation during the gauge transformation. To circumvent the high computation cost in gauge learning with regularization, we directly derive an information-invariant gauge transformation which allows to preserve scene information inherently and yield superior performance.

Despite the success of Transformers in various applications from text, vision, and speech domains, they are yet to become standard architectures for mobile and edge device applications due to their heavy memory and computational requirements. While there exist many different approaches to reduce the complexities of the Transformers, such as the pruning of the weights/attentions/tokens, quantization, and distillation, we focus on token pruning, which reduces not only the complexity of the attention operations, but also the linear layers, which have non-negligible computational costs. However, previous token pruning approaches often remove tokens during the feed-forward stage without consideration of their impact on later layers' attentions, which has a potential risk of dropping out important tokens for the given task. To tackle this issue, we propose an attention back-tracking method that tracks the importance of each attention in a Transformer architecture from the outputs to the inputs, to preserve the tokens that have a large impact on the final predictions. We experimentally validate the effectiveness of the method on both NLP and CV benchmarks, using Transformer architectures for both domains, and the results show that the proposed attention back-tracking allows the model to better retain the full models' performance even at high sparsity rates, significantly outperforming all baselines. Qualitative analysis of the examples further shows that our method does preserve semantically meaningful tokens.

Transformers have seen growing interest in processing different modalities, including language and image data. As a result, we can process vision and language data using transformers that are architecturally similar. Leveraging this feature of transformers, we propose weight sharing across two transformer backbones and within the same transformer backbone and pruning across two backbones in a unified framework. More specifically, we investigate weight sharing and pruning for two components of the transformers: (1) Multi-Head Attention (MSA) and (2) Feed-Forward Network (FFN) layers. To jointly perform weight sharing and pruning, we propose to use a regularization term to align model weights and the desired structure during the multimodal pre-training step. The structure vectors of sharing and pruning are generated by using a hypernetwork, which can capture complex interactions between pruning and sharing across layers and modalities. We train the hypernetwork and model weights iteratively so that the learned structure evolves along with model weights. After minimizing the proposed objective in the pre-training step, we perform weight sharing and pruning and fine-tune the compressed model on downstream tasks. Finally, we perform experiments on vision and language tasks, including Referring Expression Comprehension (REC), Visual Question Answering (VQA), and Object Detection using the state-of-the-art grounding based models: MDETR and GLIP. Our experiments show that we can compress these models by $35-40\%$ by sharing and pruning MSA and FFN weights without almost any loss in accuracy.

All existing 3D-from-2D generators are designed for well-curated single-category datasets, where all the objects have (approximately) the same scale, 3D location, and orientation, and the camera always points to the center of the scene. This makes them inapplicable to diverse, in-the-wild datasets of non-alignable scenes rendered from arbitrary camera poses. In this work, we develop a 3D generator with Generic Priors (3DGP): a 3D synthesis framework with more general assumptions about the training data, and show that it scales to very challenging datasets, like ImageNet. Our model is based on three new ideas. First, we incorporate an inaccurate off-the-shelf depth estimator into 3D GAN training via a special depth adaptation module to handle the imprecision. Then, we create a flexible camera model and a regularization strategy for it to learn its distribution parameters during training. Finally, we extend the recent ideas of transferring knowledge from pretrained classifiers into GANs for patch-wise trained models by employing a simple distillation-based technique on top of the discriminator. It achieves more stable training than the existing methods and speeds up the convergence by at least 40%. We explore our model on four datasets: SDIP Dogs $256^2$, SDIP Elephants $256^2$, LSUN Horses $256^2$, and ImageNet $256^2$ and demonstrate that 3DGP outperforms the recent state-of-the-art in terms of both texture and geometry quality. Code and visualizations: https://snap-research.github.io/3dgp.

We present Analogical Networks, a model that segments 3D object scenes with analogical reasoning: instead of mapping a scene to part segments directly, our model first retrieves related scenes from memory and their corresponding part structures, and then predicts analogous part structures in the input object 3D point cloud, via an end-to-end learnable modulation mechanism. By conditioning on more than one retrieved memories, compositions of structures are predicted, that mix and match parts across the retrieved memories. One-shot, few-shot or many-shot learning are treated uniformly in Analogical Networks, by conditioning on the appropriate set of memories, whether taken from a single, few or many memory exemplars, and inferring analogous parses. We show Analogical Networks are competitive with state-of-the-art 3D segmentation transformer in many-shot settings and outperform them and existing paradigms of meta-learning and few-shot learning in few-shot scenarios. Our model successfully parses instances of novel object categories simply by expanding its memory, without any weight updates.

We propose a method for 3D shape reconstruction from unoriented point clouds. Our method consists of a novel SE(3)-equivariant coordinate-based network (TF-ONet), that parametrizes the occupancy field of the shape and respects the inherent symmetries of the problem. In contrast to previous shape reconstruction methods that align the input to a regular grid, we operate directly on the irregular point cloud. Our architecture leverages equivariant attention layers that operate on local tokens. This mechanism enables local shape modelling, a crucial property for scalability to large scenes. Given an unoriented, sparse, noisy point cloud as input, we produce equivariant features for each point. These serve as keys and values for the subsequent equivariant cross-attention blocks that parametrize the occupancy field. By querying an arbitrary point in space, we predict its occupancy score. We show that our method outperforms previous SO(3)-equivariant methods, as well as non-equivariant methods trained on SO(3)-augmented datasets. More importantly, local modelling together with SE(3)-equivariance create an ideal setting for SE(3) scene reconstruction. We show that by training only on single, aligned objects and without any pre-segmentation, we can reconstruct novel scenes containing arbitrarily many objects in random poses without any performance loss.

Unsupervised contrastive learning for indoor-scene point clouds has achieved great successes. However, unsupervised representation learning on outdoor-scene point clouds remains challenging because previous methods need to reconstruct the whole scene and capture partial views for the contrastive objective. This is infeasible in outdoor scenes with moving objects, obstacles, and sensors. In this paper, we propose CO3, namely {Co}operative {Co}ntrastive Learning and {Co}ntextual Shape Prediction, to learn 3D representation for outdoor-scene point clouds in an unsupervised manner. CO3 has several merits compared to existing methods. (1) It utilizes LiDAR point clouds from vehicle-side and infrastructure-side to build views that differ enough but meanwhile maintain common semantic information for contrastive learning, which are more appropriate than views built by previous methods. (2) Alongside the contrastive objective, we propose contextual shape prediction to bring more task-relevant information for unsupervised 3D point cloud representation learning and we also provide a theoretical analysis for this pre-training goal. (3) As compared to previous methods, representation learned by CO3 is able to be transferred to different outdoor scene dataset collected by different type of LiDAR sensors. (4) CO3 improves current state-of-the-art methods on Once, KITTI and NuScenes datasets by up to 2.58 mAP in 3D object detection task and 3.54 mIoU in LiDAR semantic segmentation task. Codes and models will be released.

We present a method to map 2D image observations of a scene to a persistent 3D scene representation, enabling novel view synthesis and disentangled representation of the movable and immovable components of the scene. Motivated by the bird’s-eye-view (BEV) representation commonly used in vision and robotics, we propose conditional neural groundplans, ground-aligned 2D feature grids, as persistent and memory-efficient scene representations. Our method is trained self-supervised from unlabeled multi-view observations using differentiable rendering, and learns to complete geometry and appearance of occluded regions. In addition, we show that we can leverage multi-view videos at training time to learn to separately reconstruct static and movable components of the scene from a single image at test time. The ability to separately reconstruct movable objects enables a variety of downstream tasks using simple heuristics, such as extraction of object-centric 3D representations, novel view synthesis, instance-level segmentation, 3D bounding box prediction, and scene editing. This highlights the value of neural groundplans as a backbone for efficient 3D scene understanding models.

The Neural Radiance Fields (NeRF) have been recently applied to reconstruct building-scale and even city-scale scenes. To model a large-scale scene efficiently, a dominant strategy is to employ a divide-and-conquer paradigm via performing scene decomposition, which decomposes a complex scene into parts that are further processed by different sub-networks. Existing large-scale NeRFs mainly use heuristic hand-crafted scene decomposition, with regular 3D-distance-based or physical-street-block-based schemes. Although achieving promising results, the hand-crafted schemes limit the capabilities of NeRF in large-scale scene modeling in several aspects. Manually designing a universal scene decomposition rule for different complex scenes is challenging, leading to adaptation issues for different scenarios. The decomposition procedure is not learnable, hindering the network from jointly optimizing the scene decomposition and the radiance fields in an end-to-end manner. The different sub-networks are typically optimized independently, and thus hand-crafted rules are required to composite them to achieve a better consistency. To tackle these issues, we propose Switch-NeRF, a novel end-to-end large-scale NeRF with learning-based scene decomposition. We design a gating network to dispatch 3D points to different NeRF sub-networks. The gating network can be optimized together with the NeRF sub-networks for different scene partitions, by a design with the Sparsely Gated Mixture of Experts (MoE). The outputs from different sub-networks can also be fused in a learnable way in the unified framework to effectively guarantee the consistency of the whole scene. Furthermore, the proposed MoE-based Switch-NeRF model is carefully implemented and optimized to achieve both high-fidelity scene reconstruction and efficient computation. Our method establishes clear state-of-the-art performances on several large-scale datasets. To the best of our knowledge, we are the first to propose an applicable end-to-end sparse NeRF network with learning-based decomposition for large-scale scenes. Codes are released at https://github.com/MiZhenxing/Switch-NeRF.

Multi-view projection methods have demonstrated promising performance on 3D understanding tasks like 3D classification and segmentation. However, it remains unclear how to combine such multi-view methods with the widely available 3D point clouds. Previous methods use unlearned heuristics to combine features at the point level. To this end, we introduce the concept of the multi-view point cloud (Voint cloud), representing each 3D point as a set of features extracted from several view-points. This novel 3D Voint cloud representation combines the compactness of 3D point cloud representation with the natural view-awareness of multi-view representation. Naturally, we can equip this new representation with convolutional and pooling operations. We deploy a Voint neural network (VointNet) to learn representations in the Voint space. Our novel representation achieves state-of-the-art performance on 3D classification, shape retrieval, and robust 3D part segmentation on standard benchmarks ( ScanObjectNN, ShapeNet Core55, and ShapeNet Parts). Further analysis shows that VointNet improves the robustness to occlusion compared to other methods.

We present the Group Propagation Vision Transformer (GPViT): a novel non- hierarchical (i.e. non-pyramidal) transformer model designed for general visual recognition with high-resolution features. High-resolution features (or tokens) are a natural fit for tasks that involve perceiving fine-grained details such as detection and segmentation, but exchanging global information between these features is expensive in memory and computation because of the way self-attention scales. We provide a highly efficient alternative Group Propagation Block (GP Block) to exchange global information. In each GP Block, features are first grouped to- gether by a fixed number of learnable group tokens; we then perform Group Propagation where global information is exchanged between the grouped fea- tures; finally, global information in the updated grouped features is returned back to the image features through a transformer decoder. We evaluate GPViT on a variety of visual recognition tasks including image classification, semantic seg- mentation, object detection, and instance segmentation. Our method achieves significant performance gains over previous works across all tasks, especially on tasks that require high-resolution outputs, for example, our GPViT-L3 out- performs Swin Transformer-B by 2.0 mIoU on ADE20K semantic segmentation with only half as many parameters. Code and pre-trained models are available at https://github.com/ChenhongyiYang/GPViT.

We propose a new task to benchmark scene understanding of embodied agents: Situated Question Answering in 3D Scenes (SQA3D). Given a scene context (e.g., 3D scan), SQA3D requires the tested agent to first understand its situation (position, orientation, etc.) in the 3D scene as described by text, then reason about its surrounding environment and answer a question under that situation. Based upon 650 scenes from ScanNet, we provide a dataset centered around 6.8k unique situations, along with 20.4k descriptions and 33.4k diverse reasoning questions for these situations. These questions examine a wide spectrum of reasoning capabilities for an intelligent agent, ranging from spatial relation comprehension to commonsense understanding, navigation, and multi-hop reasoning. SQA3D imposes a significant challenge to current multi-modal especially 3D reasoning models. We evaluate various state-of-the-art approaches and find that the best one only achieves an overall score of 47.20%, while amateur human participants can reach 90.06%. We believe SQA3D could facilitate future embodied AI research with stronger situation understanding and reasoning capability.

We present 3DiM (pronounced "three-dim"), a diffusion model for 3D novel view synthesis from as few as a single image. The core of 3DiM is an image-to-image diffusion model -- 3DiM takes a single reference view and their poses as inputs, and generates a novel view via diffusion. 3DiM can then generate a full 3D consistent scene following our novel stochastic conditioning sampler: the output frames of the scene are generated autoregressively, and during the reverse diffusion process of each individual frame, we select a random conditioning frame from the set of previous frames at each denoising step. We demonstrate that stochastic conditioning yields much more 3D consistent results compared to the naive sampling process which only conditions on a single previous frame. We compare 3DiMs to prior work on the SRN ShapeNet dataset, demonstrating that 3DiM's generated videos from a single view achieve much higher fidelity while being approximately 3D consistent. We also introduce a new evaluation methodology, 3D consistency scoring, to measure the 3D consistency of a generated object by training a neural field on the model's output views. 3DiMs are geometry free, do not rely on hyper-networks or test-time optimization for novel view synthesis, and allow a single model to easily scale to a large number of scenes.

While recent camera-only 3D detection methods leverage multiple timesteps, the limited history they use significantly hampers the extent to which temporal fusion can improve object perception. Observing that existing works' fusion of multi-frame images are instances of temporal stereo matching, we find that performance is hindered by the interplay between 1) the low granularity of matching resolution and 2) the sub-optimal multi-view setup produced by limited history usage. Our theoretical and empirical analysis demonstrates that the optimal temporal difference between views varies significantly for different pixels and depths, making it necessary to fuse many timesteps over long-term history. Building on our investigation, we propose to generate a cost volume from a long history of image observations, compensating for the coarse but efficient matching resolution with a more optimal multi-view matching setup. Further, we augment the per-frame monocular depth predictions used for long-term, coarse matching with short-term, fine-grained matching and find that long and short term temporal fusion are highly complementary. While maintaining high efficiency, our framework sets new state-of-the-art on nuScenes, achieving first place on the test set and outperforming previous best art by 5.2% mAP and 3.7% NDS on the validation set. Code will be released here: https://github.com/Divadi/SOLOFusion.

This paper presents MOAT, a family of neural networks that build on top of MObile convolution (i.e., inverted residual blocks) and ATtention. Unlike the current works that stack separate mobile convolution and transformer blocks, we effectively merge them into a MOAT block. Starting with a standard Transformer block, we replace its multi-layer perceptron with a mobile convolution block, and further reorder it before the self-attention operation. The mobile convolution block not only enhances the network representation capacity, but also produces better downsampled features. Our conceptually simple MOAT networks are surprisingly effective, achieving 89.1% / 81.5% top-1 accuracy on ImageNet-1K / ImageNet-1K-V2 with ImageNet-22K pretraining. Additionally, MOAT can be seamlessly applied to downstream tasks that require large resolution inputs by simply converting the global attention to window attention. Thanks to the mobile convolution that effectively exchanges local information between pixels (and thus cross-windows), MOAT does not need the extra window-shifting mechanism. As a result, on COCO object detection, MOAT achieves 59.2% AP$^{\text{box}}$ with 227M model parameters (single-scale inference, and hard NMS), and on ADE20K semantic segmentation, MOAT attains 57.6% mIoU with 496M model parameters (single-scale inference). Finally, the tiny-MOAT family, obtained by simply reducing the channel sizes, also surprisingly outperforms several mobile-specific transformer-based models on ImageNet. The tiny-MOAT family is also benchmarked on downstream tasks, serving as a baseline for the community. We hope our simple yet effective MOAT will inspire more seamless integration of convolution and self-attention. Code is publicly available.

Text-guided 3D shape generation remains challenging due to the absence of large paired text-shape dataset, the substantial semantic gap between these two modalities, and the structural complexity of 3D shapes. This paper presents a new framework called Image as Stepping Stone (ISS) for the task by introducing 2D image as a stepping stone to connect the two modalities and to eliminate the need for paired text-shape data. Our key contribution is a two-stage feature-space-alignment approach that maps CLIP features to shapes by harnessing a pre-trained single-view reconstruction (SVR) model with multi-view supervisions: first map the CLIP image feature to the detail-rich shape space in the SVR model, then map the CLIP text feature to the shape space and optimize the mapping by encouraging CLIP consistency between the input text and the rendered images. Further, we formulate a textguided shape stylization module to dress up the output shapes with novel structures and textures. Beyond existing works on 3D shape generation from text, our new approach is general for creating shapes in a broad range of categories, without requiring paired text-shape data. Experimental results manifest that our approach outperforms the state-of-the-arts and our baselines in terms of fidelity and consistency with text. Further, our approach can stylize the generated shapes with both realistic and fantasy structures and textures. Codes are available at https://github.com/liuzhengzhe/ISS-Image-as-Stepping-Stone-for-Text-Guided-3D-Shape-Generation.

Recent breakthroughs in text-to-image synthesis have been driven by diffusion models trained on billions of image-text pairs. Adapting this approach to 3D synthesis would require large-scale datasets of labeled 3D or multiview data and efficient architectures for denoising 3D data, neither of which currently exist. In this work, we circumvent these limitations by using a pretrained 2D text-to-image diffusion model to perform text-to-3D synthesis. We introduce a loss based on probability density distillation that enables the use of a 2D diffusion model as a prior for optimization of a parametric image generator. Using this loss in a DeepDream-like procedure, we optimize a randomly-initialized 3D model (a Neural Radiance Field, or NeRF) via gradient descent such that its 2D renderings from random angles achieve a low loss. The resulting 3D model of the given text can be viewed from any angle, relit by arbitrary illumination, or composited into any 3D environment. Our approach requires no 3D training data and no modifications to the image diffusion model, demonstrating the effectiveness of pretrained image diffusion models as priors.