ICCV 2023 Papers

Recently, transformer-based methods have dominated 3D instance segmentation, where mask attention is commonly involved. Specifically, object queries are guided by the initial instance masks in the first cross-attention, and then iteratively refine themselves in a similar manner. However, we observe that the mask-attention pipeline usually leads to slow convergence due to low-recall initial instance masks. Therefore, we abandon the mask attention design and resort to an auxiliary center regression task instead. Through center regression, we effectively overcome the low-recall issue and perform cross-attention by imposing positional prior. To reach this goal, we develop a series of position-aware designs. First, we learn a spatial distribution of 3D locations as the initial position queries. They spread over the 3D space densely, and thus can easily capture the objects in a scene with a high recall. Moreover, we present relative position encoding for the cross-attention and iterative refinement for more accurate position queries. Experiments show that our approach converges 4x faster than existing work, sets a new state of the art on ScanNetv2 3D instance segmentation benchmark, and also demonstrates superior performance across various datasets. Code and models are available at https://github.com/dvlab-research/Mask-Attention-Free-Transformer.

Despite the success in large-scale text-to-image generation and text-conditioned image editing, existing methods still struggle to produce consistent generation and editing results. For example, generation approaches usually fail to synthesize multiple images of the same objects/characters but with different views or poses. Meanwhile, existing editing methods either fail to achieve effective complex non-rigid editing while maintaining the overall textures and identity, or require time-consuming fine-tuning to capture the image-specific appearance. In this paper, we develop MasaCtrl, a tuning-free method to achieve consistent image generation and complex non-rigid image editing simultaneously. Specifically, MasaCtrl converts existing self-attention in diffusion models into mutual self-attention, so that it can query correlated local contents and textures from source images for consistency. To further alleviate the query confusion between foreground and background, we propose a mask-guided mutual self-attention strategy, where the mask can be easily extracted from the cross-attention maps. Extensive experiments show that the proposed MasaCtrl can produce impressive results in both consistent image generation and complex non-rigid real image editing.

We present a new method for open-vocabulary universal image segmentation, which is capable of performing instance, semantic, and panoptic segmentation under a unified framework. Our approach, called MasQCLIP, seamlessly integrates with a pre-trained CLIP model by utilizing its dense features, thereby circumventing the need for extensive parameter training. MasQCLIP emphasizes two new aspects when building an image segmentation method with a CLIP model: 1) a student-teacher module to deal with masks of the novel (unseen) classes by distilling information from the base (seen) classes; 2) a fine-tuning process to update model parameters for the queries Q within the CLIP model. Thanks to these two simple and intuitive designs, MasQCLIP is able to achieve state-of-the-art performances with a substantial gain over the competing methods by a large margin across all three tasks, including open-vocabulary instance, semantic, and panoptic segmentation. Project page is at https://masqclip.github.io/.

This paper addresses data augmentation for action segmentation. Our key novelty is that we augment the original training videos in the deep feature space, not in the visual spatiotemporal domain as done by previous work. For augmentation, we modify original deep features of video frames such that the resulting embeddings fall closer to the class decision boundaries. Also, we edit action sequences of the original training videos (a.k.a. transcripts) by inserting, deleting, and replacing actions such that the resulting transcripts are close in edit distance to the ground truth ones. For our data augmentation we resort to reinforcement learning, instead of more common supervised learning, since we do not have access to reliable oracles which would provide supervision about the optimal data modifications in the deep feature space. For modifying frame embeddings, we use a meta-model formulated as a Markov Game with multiple self-interested agents. Also, new transcripts are generated using a fast, parameter-free Monte Carlo tree search. Our experiments show that the proposed data augmentation of the Breakfast, GTEA, and 50Salads datasets leads to significant performance gains of several state of the art action segmenters.

Many Vision-and-Language Navigation (VLN) tasks have been proposed in recent years, from room-based to object-based and indoor to outdoor. The REVERIE (Remote Embodied Referring Expression) is interesting since it only provides high-level instructions to the agent, which are closer to human commands in practice. Nevertheless, this poses more challenges than other VLN tasks since it requires agents to infer a navigation plan only based on a short instruction. Large Language Models (LLMs) show great potential in robot action planning by providing proper prompts. Still, this strategy has not been explored under the REVERIE settings. There are several new challenges. For example, the LLM should be environment-aware so that the navigation plan can be adjusted based on the current visual observation. Moreover, the LLM planned actions should be adaptable to the much larger and more complex REVERIE environment. This paper proposes a March-in-Chat (MiC) model that can talk to the LLM on the fly and plan dynamically based on a newly proposed Room-and-Object Aware Scene Perceiver (ROASP). Our MiC model outperforms the previous state-of-the-art by large margins by SPL and RGSPL metrics on the REVERIE benchmark. The source code is available at https://github.com/YanyuanQiao/MiC

Despite tremendous advancements in bird's-eye view (BEV) perception, existing models fall short in generating realistic and coherent semantic map layouts, and they fail to account for uncertainties arising from partial sensor information (such as occlusion or limited coverage). In this work, we introduce MapPrior, a novel BEV perception framework that combines a traditional discriminative BEV perception model with a learned generative model for semantic map layouts. Our MapPrior delivers predictions with better accuracy, realism, and uncertainty awareness. We evaluate our model on the large-scale nuScenes benchmark. At the time of submission, MapPrior outperforms the strongest competing method, with significantly improved MMD and ECE scores in camera- and LiDAR-based BEV perception. Furthermore, our method can be used to perpetually generate layouts with unconditional sampling.

Change detection in remote sensing imagery is essential for a variety of applications such as urban planning, disaster management, and climate research. However, existing methods for identifying semantically changed areas overlook the availability of semantic information in the form of existing maps describing features of the earth's surface. In this paper, we leverage this information for change detection in bi-temporal images. We show that the simple integration of the additional information via concatenation of latent representations suffices to significantly outperform state-of-the-art change detection methods. Motivated by this observation, we propose the new task of Conditional Change Detection, where pre-change semantic information is used as input next to bi-temporal images. To fully exploit the extra information, we propose MapFormer, a novel architecture based on a multi-modal feature fusion module that allows for feature processing conditioned on the available semantic information. We further employ a supervised, cross-modal contrastive loss to guide the learning of visual representations. Our approach outperforms existing change detection methods by an absolute 11.7% and 18.4% in terms of binary change IoU on DynamicEarthNet and HRSCD, respectively. Furthermore, we demonstrate the robustness of our approach to the quality of the pre-change semantic information and the absence pre-change imagery. The code is available at https://github.com/mxbh/mapformer.

The field of visual representation learning has seen explosive growth in the past years, but its benefits in robotics have been surprisingly limited so far. Prior work uses generic visual representations as a basis to learn (task-specific) robot action policies (e.g., via behavior cloning). While the visual representations do accelerate learning, they are primarily used to encode visual observations. Thus, action information has to be derived purely from robot data, which is expensive to collect! In this work, we present a scalable alternative where the visual representations can help directly infer robot actions. We observe that vision encoders express relationships between image observations as distances (e.g., via embedding dot product) that could be used to efficiently plan robot behavior. We operationalize this insight and develop a simple algorithm for acquiring a distance function and dynamics predictor, by fine-tuning a pre-trained representation on human collected video sequences. The final method is able to substantially outperform traditional robot learning baselines (e.g., 70% success v.s. 50% for behavior cloning on pick-place) on a suite of diverse real-world manipulation tasks. It can also generalize to novel objects, without using any robot demonstrations during train time. For visualizations of the learned policies please check: https://agi-labs.github.io/manipulate-by-seeing/

In this work, we investigate the problem of creating high-fidelity 3D content from only a single image. This is inherently challenging: it essentially involves estimating the underlying 3D geometry while hallucinating unseen textures. To address this challenge, we leverage prior knowledge in a well-trained 2D diffusion model to serve as a 3D-aware supervision for 3D creation. Our proposed method, Make-It-3D, employs a two-stage optimization pipeline: the first stage optimizes a neural radiance field with constraints from the reference image and diffusion prior; the second stage builds textured point clouds from the coarse model and further enhances the textures with diffusion prior leveraging the availability of high-quality textures from the reference image. Extensive experiments show that our method achieves a clear improvement over previous works, displaying faithful reconstruction and impressive visual quality. Our method presents the first attempt to achieve high-quality 3D creation from a single image for general objects, and enables various applications such as text-to-3D creation and texture editing.

Text-guided human motion generation has drawn significant interest because of its impactful applications spanning animation and robotics. Recently, application of diffusion models for motion generation has enabled improvements in the quality of generated motions. However, existing approaches are limited by their reliance on relatively small-scale motion capture data, leading to poor performance on more diverse, in-the-wild prompts. In this paper, we introduce Make-An-Animation, a text-conditioned human motion generation model which learns more diverse poses and prompts from large-scale image-text datasets, enabling significant improvement in performance over prior works. Make-An-Animation is trained in two stages. First, we train on a curated large-scale dataset of (text, static pseudo-pose) pairs extracted from image-text datasets. Second, we fine-tune on motion capture data, adding additional layers to model the temporal dimension. Unlike prior diffusion models for motion generation, Make-An-Animation uses a U-Net architecture similar to recent text-to-video generation models. Human evaluation of motion realism and alignment with input text shows that our model reaches state-of-the-art performance on text-to-motion generation.

3D GAN inversion aims to achieve high reconstruction fidelity and reasonable 3D geometry simultaneously from a single image input. However, existing 3D GAN inversion methods rely on time-consuming optimization for each individual case. In this work, we introduce a novel encoder-based inversion framework based on EG3D, one of the most widely-used 3D GAN models. We leverage the inherent properties of EG3D's latent space to design a discriminator and a background depth regularization. This enables us to train a geometry-aware encoder capable of converting the input image into corresponding latent code. Additionally, we explore the feature space of EG3D and develop an adaptive refinement stage that improves the representation ability of features in EG3D to enhance the recovery of fine-grained textural details. Finally, we propose an occlusion-aware fusion operation to prevent distortion in unobserved regions. Our method achieves impressive results comparable to optimization-based methods while operating up to 500 times faster. Our framework is well-suited for applications such as semantic editing.

The advent of open-source AI communities has produced a cornucopia of powerful text-guided diffusion models that are trained on various datasets. While few explorations have been conducted on ensembling such models to combine their strengths. In this work, we propose a simple yet effective method called Saliency-aware Noise Blending (SNB) that can empower the fused text-guided diffusion models to achieve more controllable generation. Specifically, we experimentally find that the responses of classifier-free guidance are highly related to the saliency of generated images. Thus we propose to trust different models in their areas of expertise by blending the predicted noises of two diffusion models in a saliency-aware manner. SNB is training-free and can be completed within a DDIM sampling process. Additionally, it can automatically align the semantics of two noise spaces without requiring additional annotations such as masks. Extensive experiments show the impressive effectiveness of SNB in various applications. The project page is available at https://magicfusion.github.io.

We propose a fast and generalizable solution to Multiview Photometric Stereo (MVPS), called MVPSNet. The key to our approach is a feature extraction network that effectively combines images from the same view captured under multiple lighting conditions to extract geometric features from shading cues for stereo matching. We demonstrate these features, termed 'Light Aggregated Feature Maps' (LAFM), are effective for feature matching even in textureless regions, where traditional multi-view stereo methods often fail. Our method produces similar reconstruction results to PS-NeRF, a state-of-the-art MVPS method that optimizes a neural network per-scene, while being 411x faster (105 seconds vs. 12 hours) in inference. Additionally, we introduce a new synthetic dataset for MVPS, sMVPS, which is shown to be effective for training a generalizable MVPS method.

While bird's-eye-view (BEV) perception models can be useful for building high-definition maps (HD-Maps) with less human labor, their results are often unreliable and demonstrate noticeable inconsistencies in the predicted HD-Maps from different viewpoints. This is because BEV perception is typically set up in an "onboard" manner, which restricts the computation and consequently prevents algorithms from reasoning multiple views simultaneously. This paper overcomes these limitations and advocates a more practical "offboard" HD-Map generation setup that removes the computation constraints, based on the fact that HD-Maps are commonly reusable infrastructures built offline in data centers. To this end, we propose a novel offboard pipeline called MV-Map that capitalizes multi-view consistency and can handle an arbitrary number of frames with the key desgin of a "region-centric" framework. In MV-Map, the target HD-Maps are created by aggregating all the frames of onboard predictions, weighted by the confidence scores assigned by an "uncertainty network." To further enhance multi-view consistency, we augment the uncertainty network with the global 3D structure optimized by a voxelized neural radiance field (Voxel-NeRF). Extensive experiments on nuScenes show that our MV-Map significantly improves the quality of HD-Maps, further highlighting the importance of offboard methods for HD-Map generation.

Reconstructing 3D vehicles from noisy and sparse partial point clouds is of great significance to autonomous driving. Most existing 3D reconstruction methods cannot be directly applied to this problem because they are elaborately designed to deal with dense inputs with trivial noise. In this work, we propose a novel framework, dubbed MV-DeepSDF, which estimates the optimal Signed Distance Function (SDF) shape representation from multi-sweep point clouds to reconstruct vehicles in the wild. Although there have been some SDF-based implicit modeling methods, they only focus on single-view-based reconstruction, resulting in low fidelity. In contrast, we first analyze multi-sweep consistency and complementarity in the latent feature space and propose to transform the implicit space shape estimation problem into an element-to-set feature extraction problem. Then, we devise a new architecture to extract individual element-level representations and aggregate them to generate a set-level predicted latent code. This set-level latent code is an expression of the optimal 3D shape in the implicit space, and can be subsequently decoded to a continuous SDF of the vehicle. In this way, our approach learns consistent and complementary information among multi-sweeps for 3D vehicle reconstruction. We conduct thorough experiments on two real-world autonomous driving datasets (Waymo and KITTI) to demonstrate the superiority of our approach over state-of-the-art alternative methods both qualitatively and quantitatively.

Machine unlearning is an emerging task of removing the influence of selected training datapoints from a trained model upon data deletion requests, which echoes the widely enforced data regulations mandating the Right to be Forgotten. Many unlearning methods have been proposed recently, achieving significant efficiency gains over the naive baseline of retraining from scratch. However, existing methods focus exclusively on unlearning from standard training models and do not apply to adversarial training models (ATMs) despite their popularity as effective defenses against adversarial examples. During adversarial training, the training data are involved in not only an outer loop for minimizing the training loss, but also an inner loop for generating the adversarial perturbation. Such bi-level optimization greatly complicates the influence measure for the data to be deleted and renders the unlearning more challenging than standard model training with single-level optimization. This paper proposes a new approach called MUter for unlearning from ATMs. We derive a closed-form unlearning step underpinned by a total Hessian-related data influence measure, while existing methods can mis-capture the data influence associated with the indirect Hessian part. We further alleviate the computational cost by introducing a series of approximations and conversions to avoid the most computationally demanding parts of Hessian inversions. The efficiency and effectiveness of MUter have been validated through experiments on four datasets using both linear and neural network models.

Amodal Instance Segmentation (AIS) endeavors to accurately deduce complete object shapes that are partially or fully occluded. However, the inherent ill-posed nature of single-view datasets poses challenges in determining occluded shapes. A multi-view framework may help alleviate this problem, as humans often adjust their perspective when encountering occluded objects. At present, this approach has not yet been explored by existing methods and datasets. To bridge this gap, we propose a new task called Multi-view Amodal Instance Segmentation (MAIS) and introduce the MUVA dataset, the first MUlti-View AIS dataset that takes the shopping scenario as instantiation. MUVA provides comprehensive annotations, including multi-view amodal/visible segmentation masks, 3D models, and depth maps, making it the largest image-level AIS dataset in terms of both the number of images and instances. Additionally, we propose a new method for aggregating representative features across different instances and views, which demonstrates promising results in accurately predicting occluded objects from one viewpoint by leveraging information from other viewpoints. Besides, we also demonstrate that MUVA can benefit the AIS task in real-world scenarios.

Image resizing operation is a fundamental preprocessing module in modern computer vision. Throughout the deep learning revolution, researchers have overlooked the potential of alternative resizing methods beyond the commonly used resizers that are readily available, such as nearest-neighbors, bilinear, and bicubic. The key question of our interest is whether the front-end resizer affects the performance of deep vision models? In this paper, we present an extremely lightweight multilayer Laplacian resizer with only a handful of trainable parameters, dubbed MULLER resizer. MULLER has a bandpass nature in that it learns to boost details in certain frequency subbands that benefit the downstream recognition models. We show that MULLER can be easily plugged into various training pipelines, and it effectively boosts the performance of the underlying vision task with little to no extra cost. Specifically, we select a state-of-the-art vision Transformer, MaxViT, as the baseline, and show that, if trained with MULLER, MaxViT gains up to 0.6% top-1 accuracy, and meanwhile enjoys 36% inference cost saving to achieve similar top-1 accuracy on ImageNet-1k, as compared to the standard training scheme. Notably, MULLER's performance also scales with model size and training data size such as ImageNet-21k and JFT, and it is widely applicable to multiple vision tasks, including image classification, object detection and segmentation, as well as image quality assessment. The code is available at https://github.com/google-research/google-research/tree/master/muller.

Binary neural networks (BNNs) have been widely adopted to reduce the computational cost and memory storage on edge-computing devices by using one bit representation for activations and weights. However, as neural networks become wider/deeper to improve accuracy and meet practical requirements, the computational burden remains a significant challenge even on the binary version. To address these issues, this paper proposes a novel method called Minimum Spanning Tree (MST) compression that learns to compress and accelerate BNNs. The proposed architecture leverages an observation from previous works that an output channel in a binary convolution can be computed using another output channel and XNOR operations with weights that differ from the weights of the reused channel. We first construct a fully connected graph with vertices corresponding to output channels, where the distance between two vertices is the number of different values between the weight sets used for these outputs. Then, the MST of the graph with the minimum depth is proposed to reorder output calculations, aiming to reduce computational cost and latency. Moreover, we propose a new learning algorithm to reduce the total MST distance during training. Experimental results on benchmark models demonstrate that our method achieves significant compression ratios with negligible accuracy drops, making it a promising approach for resource-constrained edge-computing devices.