CVPR 2023 Papers

Recent years have witnessed significant developments in point cloud processing, including classification and segmentation. However, supervised learning approaches need a lot of well-labeled data for training, and annotation is labor- and time-intensive. Self-supervised learning, on the other hand, uses unlabeled data, and pre-trains a backbone with a pretext task to extract latent representations to be used with the downstream tasks. Compared to 2D images, self-supervised learning of 3D point clouds is under-explored. Existing models, for self-supervised learning of 3D point clouds, rely on a large number of data samples, and require significant amount of computational resources and training time. To address this issue, we propose a novel contrastive learning approach, referred to as ToThePoint. Different from traditional contrastive learning methods, which maximize agreement between features obtained from a pair of point clouds formed only with different types of augmentation, ToThePoint also maximizes the agreement between the permutation invariant features and features discarded after max pooling. We first perform self-supervised learning on the ShapeNet dataset, and then evaluate the performance of the network on different downstream tasks. In the downstream task experiments, performed on the ModelNet40, ModelNet40C, ScanobjectNN and ShapeNet-Part datasets, our proposed ToThePoint achieves competitive, if not better results compared to the state-of-the-art baselines, and does so with significantly less training time (200 times faster than baselines)

We introduce the ObjectFolder Benchmark, a benchmark suite of 10 tasks for multisensory object-centric learning, centered around object recognition, reconstruction, and manipulation with sight, sound, and touch. We also introduce the ObjectFolder Real dataset, including the multisensory measurements for 100 real-world household objects, building upon a newly designed pipeline for collecting the 3D meshes, videos, impact sounds, and tactile readings of real-world objects. For each task in the ObjectFolder Benchmark, we conduct systematic benchmarking on both the 1,000 multisensory neural objects from ObjectFolder, and the real multisensory data from ObjectFolder Real. Our results demonstrate the importance of multisensory perception and reveal the respective roles of vision, audio, and touch for different object-centric learning tasks. By publicly releasing our dataset and benchmark suite, we hope to catalyze and enable new research in multisensory object-centric learning in computer vision, robotics, and beyond. Project page: https://objectfolder.stanford.edu

Human spatial attention conveys information about theregions of visual scenes that are important for perform-ing visual tasks. Prior work has shown that the informa-tion about human attention can be leveraged to benefit var-ious supervised vision tasks. Might providing this weakform of supervision be useful for self-supervised represen-tation learning? Addressing this question requires collect-ing large datasets with human attention labels. Yet, col-lecting such large scale data is very expensive. To addressthis challenge, we construct an auxiliary teacher model topredict human attention, trained on a relatively small la-beled dataset. This teacher model allows us to generate im-age (pseudo) attention labels for ImageNet. We then traina model with a primary contrastive objective; to this stan-dard configuration, we add a simple output head trained topredict the attentional map for each image, guided by thepseudo labels from teacher model. We measure the qual-ity of learned representations by evaluating classificationperformance from the frozen learned embeddings as wellas performance on image retrieval tasks. We find that thespatial-attention maps predicted from the contrastive modeltrained with teacher guidance aligns better with human at-tention compared to vanilla contrastive models. Moreover,we find that our approach improves classification accuracyand robustness of the contrastive models on ImageNet andImageNet-C. Further, we find that model representationsbecome more useful for image retrieval task as measuredby precision-recall performance on ImageNet, ImageNet-C,CIFAR10, and CIFAR10-C datasets.

Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still cannot reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel architectural components. Most importantly, it uses two new "maps" to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even during long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and using full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code and dataset are released for research purposes.

Real-world visual search systems involve deployments on multiple platforms with different computing and storage resources. Deploying a unified model that suits the minimal-constrain platforms leads to limited accuracy. It is expected to deploy models with different capacities adapting to the resource constraints, which requires features extracted by these models to be aligned in the metric space. The method to achieve feature alignments is called "compatible learning". Existing research mainly focuses on the one-to-one compatible paradigm, which is limited in learning compatibility among multiple models. We propose a Switchable representation learning Framework with Self-Compatibility (SFSC). SFSC generates a series of compatible sub-models with different capacities through one training process. The optimization of sub-models faces gradients conflict, and we mitigate this problem from the perspective of the magnitude and direction. We adjust the priorities of sub-models dynamically through uncertainty estimation to co-optimize sub-models properly. Besides, the gradients with conflicting directions are projected to avoid mutual interference. SFSC achieves state-of-the-art performance on the evaluated datasets.

In many real scenarios, data are often divided into a handful of artificial super categories in terms of expert knowledge rather than the representations of images. Concretely, a superclass may contain massive and various raw categories, such as refuse sorting. Due to the lack of common semantic features, the existing classification techniques are intractable to recognize superclass without raw class labels, thus they suffer severe performance damage or require huge annotation costs. To narrow this gap, this paper proposes a superclass learning framework, called SuperClass Learning with Representation Enhancement(SCLRE), to recognize super categories by leveraging enhanced representation. Specifically, by exploiting the self-attention technique across the batch, SCLRE collapses the boundaries of those raw categories and enhances the representation of each superclass. On the enhanced representation space, a superclass-aware decision boundary is then reconstructed. Theoretically, we prove that by leveraging attention techniques the generalization error of SCLRE can be bounded under superclass scenarios. Experimentally, extensive results demonstrate that SCLRE outperforms the baseline and other contrastive-based methods on CIFAR-100 datasets and four high-resolution datasets.

The public model zoo containing enormous powerful pretrained model families (e.g., ResNet/DeiT) has reached an unprecedented scope than ever, which significantly contributes to the success of deep learning. As each model family consists of pretrained models with diverse scales (e.g., DeiT-Ti/S/B), it naturally arises a fundamental question of how to efficiently assemble these readily available models in a family for dynamic accuracy-efficiency trade-offs at runtime. To this end, we present Stitchable Neural Networks (SN-Net), a novel scalable and efficient framework for model deployment. It cheaply produces numerous networks with different complexity and performance trade-offs given a family of pretrained neural networks, which we call anchors. Specifically, SN-Net splits the anchors across the blocks/layers and then stitches them together with simple stitching layers to map the activations from one anchor to another. With only a few epochs of training, SN-Net effectively interpolates between the performance of anchors with varying scales. At runtime, SN-Net can instantly adapt to dynamic resource constraints by switching the stitching positions. Extensive experiments on ImageNet classification demonstrate that SN-Net can obtain on-par or even better performance than many individually trained networks while supporting diverse deployment scenarios. For example, by stitching Swin Transformers, we challenge hundreds of models in Timm model zoo with a single network. We believe this new elastic model framework can serve as a strong baseline for further research in wider communities.

Instructional videos are an important resource to learn procedural tasks from human demonstrations. However, the instruction steps in such videos are typically short and sparse, with most of the video being irrelevant to the procedure. This motivates the need to temporally localize the instruction steps in such videos, i.e. the task called key-step localization. Traditional methods for key-step localization require video-level human annotations and thus do not scale to large datasets. In this work, we tackle the problem with no human supervision and introduce StepFormer, a self-supervised model that discovers and localizes instruction steps in a video. StepFormer is a transformer decoder that attends to the video with learnable queries, and produces a sequence of slots capturing the key-steps in the video. We train our system on a large dataset of instructional videos, using their automatically-generated subtitles as the only source of supervision. In particular, we supervise our system with a sequence of text narrations using an order-aware loss function that filters out irrelevant phrases. We show that our model outperforms all previous unsupervised and weakly-supervised approaches on step detection and localization by a large margin on three challenging benchmarks. Moreover, our model demonstrates an emergent property to solve zero-shot multi-step localization and outperforms all relevant baselines at this task.

While recent methods for motion and stereo estimation recover an unprecedented amount of details, such highly detailed structures are neither adequately reflected in the data of existing benchmarks nor their evaluation methodology. Hence, we introduce Spring -- a large, high-resolution, high-detail, computer-generated benchmark for scene flow, optical flow, and stereo. Based on rendered scenes from the open-source Blender movie "Spring", it provides photo-realistic HD datasets with state-of-the-art visual effects and ground truth training data. Furthermore, we provide a website to upload, analyze and compare results. Using a novel evaluation methodology based on a super-resolved UHD ground truth, our Spring benchmark can assess the quality of fine structures and provides further detailed performance statistics on different image regions. Regarding the number of ground truth frames, Spring is 60x larger than the only scene flow benchmark, KITTI 2015, and 15x larger than the well-established MPI Sintel optical flow benchmark. Initial results for recent methods on our benchmark show that estimating fine details is indeed challenging, as their accuracy leaves significant room for improvement. The Spring benchmark and the corresponding datasets are available at http://spring-benchmark.org.

Current Deep Network (DN) visualization and interpretability methods rely heavily on data space visualizations such as scoring which dimensions of the data are responsible for their associated prediction or generating new data features or samples that best match a given DN unit or representation. In this paper, we go one step further by developing the first provably exact method for computing the geometry of a DN's mapping -- including its decision boundary -- over a specified region of the data space. By leveraging the theory of Continuous Piecewise Linear (CPWL) spline DNs, SplineCam exactly computes a DN's geometry without resorting to approximations such as sampling or architecture simplification. SplineCam applies to any DN architecture based on CPWL activation nonlinearities, including (leaky) ReLU, absolute value, maxout, and max-pooling and can also be applied to regression DNs such as implicit neural representations. Beyond decision boundary visualization and characterization, SplineCam enables one to compare architectures, measure generalizability, and sample from the decision boundary on or off the data manifold. Project website: https://bit.ly/splinecam

Training Generative adversarial networks (GANs) stably is a challenging task. The generator in GANs transform noise vectors, typically Gaussian distributed, into realistic data such as images. In this paper, we propose a novel approach for training GANs with images as inputs, but without enforcing any pairwise constraints. The intuition is that images are more structured than noise, which the generator can leverage to learn a more robust transformation. The process can be made efficient by identifying closely related datasets, or a "friendly neighborhood" of the target distribution, inspiring the moniker, Spider GAN. To define friendly neighborhoods leveraging proximity between datasets, we propose a new measure called the signed inception distance (SID), inspired by the polyharmonic kernel. We show that the Spider GAN formulation results in faster convergence, as the generator can discover correspondence even between seemingly unrelated datasets, for instance, between Tiny-ImageNet and CelebA faces. Further, we demonstrate cascading Spider GAN, where the output distribution from a pre-trained GAN generator is used as the input to the subsequent network. Effectively, transporting one distribution to another in a cascaded fashion until the target is learnt -- a new flavor of transfer learning. We demonstrate the efficacy of the Spider approach on DCGAN, conditional GAN, PGGAN, StyleGAN2 and StyleGAN3. The proposed approach achieves state-of-the-art Frechet inception distance (FID) values, with one-fifth of the training iterations, in comparison to their baseline counterparts on high-resolution small datasets such as MetFaces, Ukiyo-E Faces and AFHQ-Cats.

In recent years, neural distance functions trained via volumetric ray marching have been widely adopted for multi-view 3D reconstruction. These methods, however, apply the ray marching procedure for the entire scene volume, leading to reduced sampling efficiency and, as a result, lower reconstruction quality in the areas of high-frequency details. In this work, we address this problem via joint training of the implicit function and our new coarse sphere-based surface reconstruction. We use the coarse representation to efficiently exclude the empty volume of the scene from the volumetric ray marching procedure without additional forward passes of the neural surface network, which leads to an increased fidelity of the reconstructions compared to the base systems. We evaluate our approach by incorporating it into the training procedures of several implicit surface modeling methods and observe uniform improvements across both synthetic and real-world datasets. Our codebase can be accessed via the project page.

Camera pose estimation is a key step in standard 3D reconstruction pipelines that operates on a dense set of images of a single object or scene. However, methods for pose estimation often fail when there are only a few images available because they rely on the ability to robustly identify and match visual features between pairs of images. While these methods can work robustly with dense camera views, capturing a large set of images can be time consuming or impractical. Here, we propose Sparse-View Camera Pose Regression and Refinement (SparsePose) for recovering accurate camera poses given a sparse set of wide-baseline images (fewer than 10). The method learns to regress initial camera poses and then iteratively refine them after training on a large-scale dataset of objects (Co3D: Common Objects in 3D). SparsePose significantly outperforms conventional and learning-based baselines in recovering accurate camera rotations and translations. We also demonstrate our pipeline for high-fidelity 3D reconstruction using only 5-9 images of an object.

Diffusion models have emerged as the new state-of-the-art generative model with high quality samples, with intriguing properties such as mode coverage and high flexibility. They have also been shown to be effective inverse problem solvers, acting as the prior of the distribution, while the information of the forward model can be granted at the sampling stage. Nonetheless, as the generative process remains in the same high dimensional (i.e. identical to data dimension) space, the models have not been extended to 3D inverse problems due to the extremely high memory and computational cost. In this paper, we combine the ideas from the conventional model-based iterative reconstruction with the modern diffusion models, which leads to a highly effective method for solving 3D medical image reconstruction tasks such as sparse-view tomography, limited angle tomography, compressed sensing MRI from pre-trained 2D diffusion models. In essence, we propose to augment the 2D diffusion prior with a model-based prior in the remaining direction at test time, such that one can achieve coherent reconstructions across all dimensions. Our method can be run in a single commodity GPU, and establishes the new state-of-the-art, showing that the proposed method can perform reconstructions of high fidelity and accuracy even in the most extreme cases (e.g. 2-view 3D tomography). We further reveal that the generalization capacity of the proposed method is surprisingly high, and can be used to reconstruct volumes that are entirely different from the training dataset. Code available: https://github.com/HJ-harry/DiffusionMBIR

Existing methods mainly handle single weather types. However, the connections of different weather conditions at deep representation level are usually ignored. These connections, if used properly, can generate complementary representations for each other to make up insufficient training data, obtaining positive performance gains and better generalization. In this paper, we focus on the very correlated rain and snow to explore their connections at deep representation level. Because sub-optimal connections may cause negative effect, another issue is that if rain and snow are handled in a multi-task learning way, how to find an optimal connection strategy to simultaneously improve deraining and desnowing performance. To build desired connection, we propose a smart knowledge assignment strategy, called SmartAssign, to optimally assign the knowledge learned from both tasks to a specific one. In order to further enhance the accuracy of knowledge assignment, we propose a novel knowledge contrast mechanism, so that the knowledge assigned to different tasks preserves better uniqueness. The inherited inductive biases usually limit the modelling ability of CNNs, we introduce a novel transformer block to constitute the backbone of our network to effectively combine long-range context dependency and local image details. Extensive experiments on seven benchmark datasets verify that proposed SmartAssign explores effective connection between rain and snow, and improves the performances of both deraining and desnowing apparently. The implementation code will be available at https://gitee.com/mindspore/models/tree/master/research/cv/SmartAssign.

Dataset distillation, also known as dataset condensation, aims to compress a large dataset into a compact synthetic one. Existing methods perform dataset condensation by assuming a fixed storage or transmission budget. When the budget changes, however, they have to repeat the synthesizing process with access to original datasets, which is highly cumbersome if not infeasible at all. In this paper, we explore the problem of slimmable dataset condensation, to extract a smaller synthetic dataset given only previous condensation results. We first study the limitations of existing dataset condensation algorithms on such a successive compression setting and identify two key factors: (1) the inconsistency of neural networks over different compression times and (2) the underdetermined solution space for synthetic data. Accordingly, we propose a novel training objective for slimmable dataset condensation to explicitly account for both factors. Moreover, synthetic datasets in our method adopt an significance-aware parameterization. Theoretical derivation indicates that an upper-bounded error can be achieved by discarding the minor components without training. Alternatively, if training is allowed, this strategy can serve as a strong initialization that enables a fast convergence. Extensive comparisons and ablations demonstrate the superiority of the proposed solution over existing methods on multiple benchmarks.

Recently, a number of iterative learning methods have been introduced to improve generalization. These typically rely on training for longer periods of time in exchange for improved generalization. LLF (later-layer-forgetting) is a state-of-the-art method in this category. It strengthens learning in early layers by periodically re-initializing the last few layers of the network. Our principal innovation in this work is to use Simulated annealing in EArly Layers (SEAL) of the network in place of re-initialization of later layers. Essentially, later layers go through the normal gradient descent process, while the early layers go through short stints of gradient ascent followed by gradient descent. Extensive experiments on the popular Tiny-ImageNet dataset benchmark and a series of transfer learning and few-shot learning tasks show that we outperform LLF by a significant margin. We further show that, compared to normal training, LLF features, although improving on the target task, degrade the transfer learning performance across all datasets we explored. In comparison, our method outperforms LLF across the same target datasets by a large margin. We also show that the prediction depth of our method is significantly lower than that of LLF and normal training, indicating on average better prediction performance.

Object-centric learning (OCL) aspires general and com- positional understanding of scenes by representing a scene as a collection of object-centric representations. OCL has also been extended to multi-view image and video datasets to apply various data-driven inductive biases by utilizing geometric or temporal information in the multi-image data. Single-view images carry less information about how to disentangle a given scene than videos or multi-view im- ages do. Hence, owing to the difficulty of applying induc- tive biases, OCL for single-view images still remains chal- lenging, resulting in inconsistent learning of object-centric representation. To this end, we introduce a novel OCL framework for single-view images, SLot Attention via SHep- herding (SLASH), which consists of two simple-yet-effective modules on top of Slot Attention. The new modules, At- tention Refining Kernel (ARK) and Intermediate Point Pre- dictor and Encoder (IPPE), respectively, prevent slots from being distracted by the background noise and indicate lo- cations for slots to focus on to facilitate learning of object- centric representation. We also propose a weak- and semi- supervision approach for OCL, whilst our proposed frame- work can be used without any assistant annotation during the inference. Experiments show that our proposed method enables consistent learning of object-centric representa- tion and achieves strong performance across four datasets. Code is available at https://github.com/object- understanding/SLASH.

Editing 3D geometry is a challenging task requiring specialized skills. In this work, we aim to facilitate the task of editing the geometry of 3D models through the use of natural language. For example, we may want to modify a 3D chair model to "make its legs thinner" or to "open a hole in its back". To tackle this problem in a manner that promotes open-ended language use and enables fine-grained shape edits, we introduce the most extensive existing corpus of natural language utterances describing shape differences: ShapeTalk. ShapeTalk contains over half a million discriminative utterances produced by contrasting the shapes of common 3D objects for a variety of object classes and degrees of similarity. We also introduce a generic framework, ChangeIt3D, which builds on ShapeTalk and can use an arbitrary 3D generative model of shapes to produce edits that align the output better with the edit or deformation description. Finally, we introduce metrics for the quantitative evaluation of language-assisted shape editing methods that reflect key desiderata within this editing setup. We note that ShapeTalk allows methods to be trained with explicit 3D-to-language data, bypassing the necessity of "lifting" 2D to 3D using methods like neural rendering, as required by extant 2D image-language foundation models. Our code and data are publicly available at https://changeit3d.github.io/.