CVPR 2023 论文列表

Learning generative image models from various domains efficiently needs transferring knowledge from an image synthesis model trained on a large dataset. We present a recipe for learning vision transformers by generative knowledge transfer. We base our framework on generative vision transformers representing an image as a sequence of visual tokens with the autoregressive or non-autoregressive transformers. To adapt to a new domain, we employ prompt tuning, which prepends learnable tokens called prompts to the image token sequence and introduces a new prompt design for our task. We study on a variety of visual domains with varying amounts of training images. We show the effectiveness of knowledge transfer and a significantly better image generation quality. Code is available at https://github.com/google-research/generative_transfer.

Scale is the primary factor for building a powerful foundation model that could well generalize to a variety of downstream tasks. However, it is still challenging to train video foundation models with billions of parameters. This paper shows that video masked autoencoder (VideoMAE) is a scalable and general self-supervised pre-trainer for building video foundation models. We scale the VideoMAE in both model and data with a core design. Specifically, we present a dual masking strategy for efficient pre-training, with an encoder operating on a subset of video tokens and a decoder processing another subset of video tokens. Although VideoMAE is very efficient due to high masking ratio in encoder, masking decoder can still further reduce the overall computational cost. This enables the efficient pre-training of billion-level models in video. We also introduce a progressive training paradigm that involves initial pre-training on the diverse multi-sourced unlabeled dataset, followed by fine-tuning on a mixed labeled dataset. Finally, we successfully train a video ViT model with a billion parameters, which achieves a new state-of-the-art performance on the datasets of Kinetics (90.0% on K400 and 89.9% on K600) and Something-Something (68.7% on V1 and 77.0% on V2). In addition, we extensively verify the pre-trained video ViT models on a variety of downstream tasks, demonstrating its effectiveness as a general video representation learner.

Masked autoencoder (MAE), a simple and effective self-supervised learning framework based on the reconstruction of masked image regions, has recently achieved prominent success in a variety of vision tasks. Despite the emergence of intriguing empirical observations on MAE, a theoretically principled understanding is still lacking. In this work, we formally characterize and justify existing empirical insights and provide theoretical guarantees of MAE. We formulate the underlying data-generating process as a hierarchical latent variable model, and show that under reasonable assumptions, MAE provably identifies a set of latent variables in the hierarchical model, explaining why MAE can extract high-level information from pixels. Further, we show how key hyperparameters in MAE (the masking ratio and the patch size) determine which true latent variables to be recovered, therefore influencing the level of semantic information in the representation. Specifically, extremely large or small masking ratios inevitably lead to low-level representations. Our theory offers coherent explanations of existing empirical observations and provides insights for potential empirical improvements and fundamental limitations of the masked-reconstruction paradigm. We conduct extensive experiments to validate our theoretical insights.

Understanding how well a deep generative model captures a distribution of high-dimensional data remains an important open challenge. It is especially difficult for certain model classes, such as Generative Adversarial Networks and Diffusion Models, whose models do not admit exact likelihoods. In this work, we demonstrate that generalized empirical likelihood (GEL) methods offer a family of diagnostic tools that can identify many deficiencies of deep generative models (DGMs). We show, with appropriate specification of moment conditions, that the proposed method can identify which modes have been dropped, the degree to which DGMs are mode imbalanced, and whether DGMs sufficiently capture intra-class diversity. We show how to combine techniques from Maximum Mean Discrepancy and Generalized Empirical Likelihood to create not only distribution tests that retain per-sample interpretability, but also metrics that include label information. We find that such tests predict the degree of mode dropping and mode imbalance up to 60% better than metrics such as improved precision/recall.

With generative models proliferating at a rapid rate, there is a growing need for general purpose fake image detectors. In this work, we first show that the existing paradigm, which consists of training a deep network for real-vs-fake classification, fails to detect fake images from newer breeds of generative models when trained to detect GAN fake images. Upon analysis, we find that the resulting classifier is asymmetrically tuned to detect patterns that make an image fake. The real class becomes a 'sink' class holding anything that is not fake, including generated images from models not accessible during training. Building upon this discovery, we propose to perform real-vs-fake classification without learning; i.e., using a feature space not explicitly trained to distinguish real from fake images. We use nearest neighbor and linear probing as instantiations of this idea. When given access to the feature space of a large pretrained vision-language model, the very simple baseline of nearest neighbor classification has surprisingly good generalization ability in detecting fake images from a wide variety of generative models; e.g., it improves upon the SoTA by +15.07 mAP and +25.90% acc when tested on unseen diffusion and autoregressive models.

Transferability of adversarial examples is critical for black-box deep learning model attacks. While most existing studies focus on enhancing the transferability of untargeted adversarial attacks, few of them studied how to generate transferable targeted adversarial examples that can mislead models into predicting a specific class. Moreover, existing transferable targeted adversarial attacks usually fail to sufficiently characterize the target class distribution, thus suffering from limited transferability. In this paper, we propose the Transferable Targeted Adversarial Attack (TTAA), which can capture the distribution information of the target class from both label-wise and feature-wise perspectives, to generate highly transferable targeted adversarial examples. To this end, we design a generative adversarial training framework consisting of a generator to produce targeted adversarial examples, and feature-label dual discriminators to distinguish the generated adversarial examples from the target class images. Specifically, we design the label discriminator to guide the adversarial examples to learn label-related distribution information about the target class. Meanwhile, we design a feature discriminator, which extracts the feature-wise information with strong cross-model consistency, to enable the adversarial examples to learn the transferable distribution information. Furthermore, we introduce the random perturbation dropping to further enhance the transferability by augmenting the diversity of adversarial examples used in the training process. Experiments demonstrate that our method achieves excellent performance on the transferability of targeted adversarial examples. The targeted fooling rate reaches 95.13% when transferred from VGG-19 to DenseNet-121, which significantly outperforms the state-of-the-art methods.

Diffusion-based generative models have achieved remarkable success in image generation. Their guidance formulation allows an external model to plug-and-play control the generation process for various tasks without fine-tuning the diffusion model. However, the direct use of publicly available off-the-shelf models for guidance fails due to their poor performance on noisy inputs. For that, the existing practice is to fine-tune the guidance models with labeled data corrupted with noises. In this paper, we argue that this practice has limitations in two aspects: (1) performing on inputs with extremely various noises is too hard for a single guidance model; (2) collecting labeled datasets hinders scaling up for various tasks. To tackle the limitations, we propose a novel strategy that leverages multiple experts where each expert is specialized in a particular noise range and guides the reverse process of the diffusion at its corresponding timesteps. However, as it is infeasible to manage multiple networks and utilize labeled data, we present a practical guidance framework termed Practical Plug-And-Play (PPAP), which leverages parameter-efficient fine-tuning and data-free knowledge transfer. We exhaustively conduct ImageNet class conditional generation experiments to show that our method can successfully guide diffusion with small trainable parameters and no labeled data. Finally, we show that image classifiers, depth estimators, and semantic segmentation models can guide publicly available GLIDE through our framework in a plug-and-play manner. Our code is available at https://github.com/riiid/PPAP.

Autoregressive transformers have shown remarkable success in video generation. However, the transformers are prohibited from directly learning the long-term dependency in videos due to the quadratic complexity of self-attention, and inherently suffering from slow inference time and error propagation due to the autoregressive process. In this paper, we propose Memory-efficient Bidirectional Transformer (MeBT) for end-to-end learning of long-term dependency in videos and fast inference. Based on recent advances in bidirectional transformers, our method learns to decode the entire spatio-temporal volume of a video in parallel from partially observed patches. The proposed transformer achieves a linear time complexity in both encoding and decoding, by projecting observable context tokens into a fixed number of latent tokens and conditioning them to decode the masked tokens through the cross-attention. Empowered by linear complexity and bidirectional modeling, our method demonstrates significant improvement over the autoregressive Transformers for generating moderately long videos in both quality and speed.

Visual dialog (VisDial) is a task of answering a sequence of questions grounded in an image, using the dialog history as context. Prior work has trained the dialog agents solely on VisDial data via supervised learning or leveraged pre-training on related vision-and-language datasets. This paper presents a semi-supervised learning approach for visually-grounded dialog, called Generative Self-Training (GST), to leverage unlabeled images on the Web. Specifically, GST first retrieves in-domain images through out-of-distribution detection and generates synthetic dialogs regarding the images via multimodal conditional text generation. GST then trains a dialog agent on the synthetic and the original VisDial data. As a result, GST scales the amount of training data up to an order of magnitude that of VisDial (1.2M to 12.9M QA data). For robust training of the synthetic dialogs, we also propose perplexity-based data selection and multimodal consistency regularization. Evaluation on VisDial v1.0 and v0.9 datasets shows that GST achieves new state-of-the-art results on both datasets. We further observe the robustness of GST against both visual and textual adversarial attacks. Finally, GST yields strong performance gains in the low-data regime. Code is available at https://github.com/gicheonkang/gst-visdial.

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.

Generative models have shown great promise in synthesizing photorealistic 3D objects, but they require large amounts of training data. We introduce SinGRAF, a 3D-aware generative model that is trained with a few input images of a single scene. Once trained, SinGRAF generates different realizations of this 3D scene that preserve the appearance of the input while varying scene layout. For this purpose, we build on recent progress in 3D GAN architectures and introduce a novel progressive-scale patch discrimination approach during training. With several experiments, we demonstrate that the results produced by SinGRAF outperform the closest related works in both quality and diversity by a large margin.

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/.

What is a rose, visually? A rose comprises its intrinsics, including the distribution of geometry, texture, and material specific to its object category. With knowledge of these intrinsic properties, we may render roses of different sizes and shapes, in different poses, and under different lighting conditions. In this work, we build a generative model that learns to capture such object intrinsics from a single image, such as a photo of a bouquet. Such an image includes multiple instances of an object type. These instances all share the same intrinsics, but appear different due to a combination of variance within these intrinsics and differences in extrinsic factors, such as pose and illumination. Experiments show that our model successfully learns object intrinsics (distribution of geometry, texture, and material) for a wide range of objects, each from a single Internet image. Our method achieves superior results on multiple downstream tasks, including intrinsic image decomposition, shape and image generation, view synthesis, and relighting.

Adversarial training is widely used to make classifiers robust to a specific threat or adversary, such as l_p-norm bounded perturbations of a given p-norm. However, existing methods for training classifiers robust to multiple threats require knowledge of all attacks during training and remain vulnerable to unseen distribution shifts. In this work, we describe how to obtain adversarially-robust model soups (i.e., linear combinations of parameters) that smoothly trade-off robustness to different l_p-norm bounded adversaries. We demonstrate that such soups allow us to control the type and level of robustness, and can achieve robustness to all threats without jointly training on all of them. In some cases, the resulting model soups are more robust to a given l_p-norm adversary than the constituent model specialized against that same adversary. Finally, we show that adversarially-robust model soups can be a viable tool to adapt to distribution shifts from a few examples.

In this paper, we extend scene understanding to include that of human sketch. The result is a complete trilogy of scene representation from three diverse and complementary modalities -- sketch, photo, and text. Instead of learning a rigid three-way embedding and be done with it, we focus on learning a flexible joint embedding that fully supports the "optionality" that this complementarity brings. Our embedding supports optionality on two axis: (i) optionality across modalities -- use any combination of modalities as query for downstream tasks like retrieval, (ii) optionality across tasks -- simultaneously utilising the embedding for either discriminative (e.g., retrieval) or generative tasks (e.g., captioning). This provides flexibility to end-users by exploiting the best of each modality, therefore serving the very purpose behind our proposal of a trilogy at the first place. First, a combination of information-bottleneck and conditional invertible neural networks disentangle the modality-specific component from modality-agnostic in sketch, photo, and text. Second, the modality-agnostic instances from sketch, photo, and text are synergised using a modified cross-attention. Once learned, we show our embedding can accommodate a multi-facet of scene-related tasks, including those enabled for the first time by the inclusion of sketch, all without any task-specific modifications. Project Page: http://www.pinakinathc.me/scenetrilogy

Generating a high-quality High Dynamic Range (HDR) image from dynamic scenes has recently been extensively studied by exploiting Deep Neural Networks (DNNs). Most DNNs-based methods require a large amount of training data with ground truth, requiring tedious and time-consuming work. Few-shot HDR imaging aims to generate satisfactory images with limited data. However, it is difficult for modern DNNs to avoid overfitting when trained on only a few images. In this work, we propose a novel semi-supervised approach to realize few-shot HDR imaging via two stages of training, called SSHDR. Unlikely previous methods, directly recovering content and removing ghosts simultaneously, which is hard to achieve optimum, we first generate content of saturated regions with a self-supervised mechanism and then address ghosts via an iterative semi-supervised learning framework. Concretely, considering that saturated regions can be regarded as masking Low Dynamic Range (LDR) input regions, we design a Saturated Mask AutoEncoder (SMAE) to learn a robust feature representation and reconstruct a non-saturated HDR image. We also propose an adaptive pseudo-label selection strategy to pick high-quality HDR pseudo-labels in the second stage to avoid the effect of mislabeled samples. Experiments demonstrate that SSHDR outperforms state-of-the-art methods quantitatively and qualitatively within and across different datasets, achieving appealing HDR visualization with few labeled samples.

Contrastive learning-based vision-language pre-training approaches, such as CLIP, have demonstrated great success in many vision-language tasks. These methods achieve cross-modal alignment by encoding a matched image-text pair with similar feature embeddings, which are generated by aggregating information from visual patches and language tokens. However, direct aligning cross-modal information using such representations is challenging, as visual patches and text tokens differ in semantic levels and granularities. To alleviate this issue, we propose a Finite Discrete Tokens (FDT) based multimodal representation. FDT is a set of learnable tokens representing certain visual-semantic concepts. Both images and texts are embedded using shared FDT by first grounding multimodal inputs to FDT space and then aggregating the activated FDT representations. The matched visual and semantic concepts are enforced to be represented by the same set of discrete tokens by a sparse activation constraint. As a result, the granularity gap between the two modalities is reduced. Through both quantitative and qualitative analyses, we demonstrate that using FDT representations in CLIP-style models improves cross-modal alignment and performance in visual recognition and vision-language downstream tasks. Furthermore, we show that our method can learn more comprehensive representations, and the learned FDT capture meaningful cross-modal correspondence, ranging from objects to actions and attributes.

This work presents the restoration of drawings of wooden built heritage. Hand-drawn drawings contain the most important original information but are often severely degraded over time. A novel restoration method based on the vector quantized variational autoencoders is presented. Latent space representations of drawings and noise are learned, which are used to map noisy drawings to clean drawings for restoration and to generate authentic noisy drawings for data augmentation. The proposed method is applied to the drawings archived in the Cultural Heritage Administration. Restored drawings show significant quality improvement and allow more accurate interpretations of information.