Multi-modal image-text models such as CLIP and LiT have demonstrated impressive performance on image classification benchmarks and their zero-shot generalization ability is particularly exciting. While the top-5 zero-shot accuracies of these models are very high, the top-1 accuracies are much lower (over 25% gap in some cases). We investigate the reasons for this performance gap and find that many of the failure cases are caused by ambiguity in the text prompts. First, we develop a simple and efficient zero-shot post-hoc method to identify images whose top-1 prediction is likely to be incorrect, by measuring consistency of the predictions w.r.t. multiple prompts and image transformations. We show that our procedure better predicts mistakes, outperforming the popular max logit baseline on selective prediction tasks. Next, we propose a simple and efficient way to improve accuracy on such uncertain images by making use of the WordNet hierarchy; specifically we augment the original class by incorporating its parent and children from the semantic label hierarchy, and plug the augmentation into text promts. We conduct experiments on both CLIP and LiT models with five different ImageNet-based datasets. For CLIP, our method improves the top-1 accuracy by 17.13% on the uncertain subset and 3.6% on the entire ImageNet validation set. We also show that our method improves across ImageNet shifted datasets and other model architectures such as LiT. Our proposed method is hyperparameter-free, requires no additional model training and can be easily scaled to other large multi-modal architectures.
Vision-language foundation models pretrained on large-scale data provide a powerful tool for many visual understanding tasks. Notably, many vision-language models build two encoders (visual and textual) that can map two modalities into the same embedding space. As a result, the learned representations achieve good zero-shot performance on tasks like image classification. However, when there are only a few examples per category, the potential of large vision-language models is often underperformed, mainly due to the gap between a large number of parameters and a relatively small amount of training data. This paper shows that we can significantly improve the performance of few-shot classification by using the category names to initialize the classification head. More interestingly, we can borrow the non-perfect category names, or even names from a foreign language, to improve the few-shot classification performance compared with random initialization. With the proposed category name initialization method, our model obtains the state-of-the-art performance on a number of few-shot image classification benchmarks (e.g., 87.37\% on ImageNet and 96.08\% on Stanford Cars, both using five-shot learning). We also investigate and analyze when the benefit of category names diminishes and how to use distillation to improve the performance of smaller models, providing guidance for future research.
Generating images from graph-structured inputs, such as scene graphs, is uniquely challenging due to the difficulty of aligning nodes and connections in graphs with objects and their relations in images. Most existing methods address this challenge by using scene layouts, which are image-like representations of scene graphs designed to capture the coarse structures of scene images. Because scene layouts are manually crafted, the alignment with images may not be fully optimized, causing suboptimal compliance between the generated images and the original scene graphs. To tackle this issue, we propose to learn scene graph embeddings by directly optimizing their alignment with images. Specifically, we pre-train an encoder to extract both global and local information from scene graphs that are predictive of the corresponding images, relying on two loss functions: masked autoencoding loss and contrastive loss. The former trains embeddings by reconstructing randomly masked image regions, while the latter trains embeddings to discriminate between compliant and non-compliant images according to the scene graph. Given these embeddings, we build a latent diffusion model to generate images from scene graphs. The resulting method, called SGDiff, allows for the semantic manipulation of generated images by modifying scene graph nodes and connections. On the Visual Genome and COCO-Stuff datasets, we demonstrate that SGDiff outperforms state-of-the-art methods, as measured by both the Inception Score and Fr\'echet Inception Distance (FID) metrics. We will release our source code and trained models at https://github.com/YangLing0818/SGDiff.
In dense image segmentation tasks (e.g., semantic, panoptic), existing methods can hardly generalize well to unseen image domains, predefined classes, and image resolution & quality variations. Motivated by these observations, we construct a large-scale entity segmentation dataset to explore fine-grained entity segmentation, with a strong focus on open-world and high-quality dense segmentation. The dataset contains images spanning diverse image domains and resolutions, along with high-quality mask annotations for training and testing. Given the high-quality and -resolution nature of the dataset, we propose CropFormer for high-quality segmentation, which can improve mask prediction using high-res image crops that provide more fine-grained image details than the full image. CropFormer is the first query-based Transformer architecture that can effectively ensemble mask predictions from multiple image crops, by learning queries that can associate the same entities across the full image and its crop. With CropFormer, we achieve a significant AP gain of $1.9$ on the challenging fine-grained entity segmentation task. The dataset and code will be released at http://luqi.info/entityv2.github.io/.
Generative Adversarial Networks (GANs) typically suffer from overfitting when limited training data is available. To facilitate GAN training, current methods propose to use data-specific augmentation techniques. Despite the effectiveness, it is difficult for these methods to scale to practical applications. In this work, we present ScoreMix, a novel and scalable data augmentation approach for various image synthesis tasks. We first produce augmented samples using the convex combinations of the real samples. Then, we optimize the augmented samples by minimizing the norms of the data scores, i.e., the gradients of the log-density functions. This procedure enforces the augmented samples close to the data manifold. To estimate the scores, we train a deep estimation network with multi-scale score matching. For different image synthesis tasks, we train the score estimation network using different data. We do not require the tuning of the hyperparameters or modifications to the network architecture. The ScoreMix method effectively increases the diversity of data and reduces the overfitting problem. Moreover, it can be easily incorporated into existing GAN models with minor modifications. Experimental results on numerous tasks demonstrate that GAN models equipped with the ScoreMix method achieve significant improvements.
Facial Attribute Manipulation (FAM) aims to aesthetically modify a given face image to render desired attributes, which has received significant attention due to its broad practical applications ranging from digital entertainment to biometric forensics. In the last decade, with the remarkable success of Generative Adversarial Networks (GANs) in synthesizing realistic images, numerous GAN-based models have been proposed to solve FAM with various problem formulation approaches and guiding information representations. This paper presents a comprehensive survey of GAN-based FAM methods with a focus on summarizing their principal motivations and technical details. The main contents of this survey include: (i) an introduction to the research background and basic concepts related to FAM, (ii) a systematic review of GAN-based FAM methods in three main categories, and (iii) an in-depth discussion of important properties of FAM methods, open issues, and future research directions. This survey not only builds a good starting point for researchers new to this field but also serves as a reference for the vision community.
Diffusion models are a class of deep generative models that have shown impressive results on various tasks with a solid theoretical foundation. Despite demonstrated success than state-of-the-art approaches, diffusion models often entail costly sampling procedures and sub-optimal likelihood estimation. Significant efforts have been made to improve the performance of diffusion models in various aspects. In this article, we present a comprehensive review of existing variants of diffusion models. Specifically, we provide the taxonomy of diffusion models and categorize them into three types: sampling-acceleration enhancement, likelihood-maximization enhancement, and data-generalization enhancement. We also introduce the other generative models (i.e., variational autoencoders, generative adversarial networks, normalizing flow, autoregressive models, and energy-based models) and discuss the connections between diffusion models and these generative models. Then we review the applications of diffusion models, including computer vision, natural language processing, waveform signal processing, multi-modal modeling, molecular graph generation, time series modeling, and adversarial purification. Furthermore, we propose new perspectives pertaining to the development of generative models. Github: https://github.com/YangLing0818/Diffusion-Models-Papers-Survey-Taxonomy.
Estimating 3D human pose and shape from 2D images is a crucial yet challenging task. While prior methods with model-based representations can perform reasonably well on whole-body images, they often fail when parts of the body are occluded or outside the frame. Moreover, these results usually do not faithfully capture the human silhouettes due to their limited representation power of deformable models (e.g., representing only the naked body). An alternative approach is to estimate dense vertices of a predefined template body in the image space. Such representations are effective in localizing vertices within an image but cannot handle out-of-frame body parts. In this work, we learn dense human body estimation that is robust to partial observations. We explicitly model the visibility of human joints and vertices in the x, y, and z axes separately. The visibility in x and y axes help distinguishing out-of-frame cases, and the visibility in depth axis corresponds to occlusions (either self-occlusions or occlusions by other objects). We obtain pseudo ground-truths of visibility labels from dense UV correspondences and train a neural network to predict visibility along with 3D coordinates. We show that visibility can serve as 1) an additional signal to resolve depth ordering ambiguities of self-occluded vertices and 2) a regularization term when fitting a human body model to the predictions. Extensive experiments on multiple 3D human datasets demonstrate that visibility modeling significantly improves the accuracy of human body estimation, especially for partial-body cases. Our project page with code is at: https://github.com/chhankyao/visdb.