We study the task of extending the large language model (LLM) into a vision-language instruction-following model. This task is crucial but challenging since the LLM is trained on text modality only, making it hard to effectively digest the visual modality. To address this, existing methods typically train a visual adapter to align the representation between a pre-trained vision transformer (ViT) and the LLM by a generative image captioning loss. However, we find that the generative objective can only produce weak alignment for vision and language, making the aligned vision-language model very hungry for the instruction fine-tuning data. In this paper, we propose CG-VLM that applies both Contrastive and Generative alignment objectives to effectively align the representation of ViT and LLM. Different from image level and sentence level alignment in common contrastive learning settings, CG-VLM aligns the image-patch level features and text-token level embeddings, which, however, is very hard to achieve as no explicit grounding patch-token relation provided in standard image captioning datasets. To address this issue, we propose to maximize the averaged similarity between pooled image-patch features and text-token embeddings. Extensive experiments demonstrate that the proposed CG-VLM produces strong vision-language alignment and is an efficient instruction learner. For example, using only 10% instruction tuning data, we reach 95% performance of state-of-the-art method LLaVA [29] on the zero-shot ScienceQA-Image benchmark.
Image super-resolution (SR) aims to learn a mapping from low-resolution (LR) to high-resolution (HR) using paired HR-LR training images. Conventional SR methods typically gather the paired training data by synthesizing LR images from HR images using a predetermined degradation model, e.g., Bicubic down-sampling. However, the realistic degradation type of test images may mismatch with the training-time degradation type due to the dynamic changes of the real-world scenarios, resulting in inferior-quality SR images. To address this, existing methods attempt to estimate the degradation model and train an image-specific model, which, however, is quite time-consuming and impracticable to handle rapidly changing domain shifts. Moreover, these methods largely concentrate on the estimation of one degradation type (e.g., blur degradation), overlooking other degradation types like noise and JPEG in real-world test-time scenarios, thus limiting their practicality. To tackle these problems, we present an efficient test-time adaptation framework for SR, named SRTTA, which is able to quickly adapt SR models to test domains with different/unknown degradation types. Specifically, we design a second-order degradation scheme to construct paired data based on the degradation type of the test image, which is predicted by a pre-trained degradation classifier. Then, we adapt the SR model by implementing feature-level reconstruction learning from the initial test image to its second-order degraded counterparts, which helps the SR model generate plausible HR images. Extensive experiments are conducted on newly synthesized corrupted DIV2K datasets with 8 different degradations and several real-world datasets, demonstrating that our SRTTA framework achieves an impressive improvement over existing methods with satisfying speed. The source code is available at https://github.com/DengZeshuai/SRTTA.
Learning to navigate to an image-specified goal is an important but challenging task for autonomous systems. The agent is required to reason the goal location from where a picture is shot. Existing methods try to solve this problem by learning a navigation policy, which captures semantic features of the goal image and observation image independently and lastly fuses them for predicting a sequence of navigation actions. However, these methods suffer from two major limitations. 1) They may miss detailed information in the goal image, and thus fail to reason the goal location. 2) More critically, it is hard to focus on the goal-relevant regions in the observation image, because they attempt to understand observation without goal conditioning. In this paper, we aim to overcome these limitations by designing a Fine-grained Goal Prompting (FGPrompt) method for image-goal navigation. In particular, we leverage fine-grained and high-resolution feature maps in the goal image as prompts to perform conditioned embedding, which preserves detailed information in the goal image and guides the observation encoder to pay attention to goal-relevant regions. Compared with existing methods on the image-goal navigation benchmark, our method brings significant performance improvement on 3 benchmark datasets (i.e., Gibson, MP3D, and HM3D). Especially on Gibson, we surpass the state-of-the-art success rate by 8% with only 1/50 model size. Project page: https://xinyusun.github.io/fgprompt-pages
Text-to-image synthesis has made encouraging progress and attracted lots of public attention recently. However, popular evaluation metrics in this area, like the Inception Score and Fr'echet Inception Distance, incur several issues. First of all, they cannot explicitly assess the perceptual quality of generated images and poorly reflect the semantic alignment of each text-image pair. Also, they are inefficient and need to sample thousands of images to stabilise their evaluation results. In this paper, we propose to evaluate text-to-image generation performance by directly estimating the likelihood of the generated images using a pre-trained likelihood-based text-to-image generative model, i.e., a higher likelihood indicates better perceptual quality and better text-image alignment. To prevent the likelihood of being dominated by the non-crucial part of the generated image, we propose several new designs to develop a credit assignment strategy based on the semantic and perceptual significance of the image patches. In the experiments, we evaluate the proposed metric on multiple popular text-to-image generation models and datasets in accessing both the perceptual quality and the text-image alignment. Moreover, it can successfully assess the generation ability of these models with as few as a hundred samples, making it very efficient in practice.
We study the task of zero-shot vision-and-language navigation (ZS-VLN), a practical yet challenging problem in which an agent learns to navigate following a path described by language instructions without requiring any path-instruction annotation data. Normally, the instructions have complex grammatical structures and often contain various action descriptions (e.g., "proceed beyond", "depart from"). How to correctly understand and execute these action demands is a critical problem, and the absence of annotated data makes it even more challenging. Note that a well-educated human being can easily understand path instructions without the need for any special training. In this paper, we propose an action-aware zero-shot VLN method ($A^2$Nav) by exploiting the vision-and-language ability of foundation models. Specifically, the proposed method consists of an instruction parser and an action-aware navigation policy. The instruction parser utilizes the advanced reasoning ability of large language models (e.g., GPT-3) to decompose complex navigation instructions into a sequence of action-specific object navigation sub-tasks. Each sub-task requires the agent to localize the object and navigate to a specific goal position according to the associated action demand. To accomplish these sub-tasks, an action-aware navigation policy is learned from freely collected action-specific datasets that reveal distinct characteristics of each action demand. We use the learned navigation policy for executing sub-tasks sequentially to follow the navigation instruction. Extensive experiments show $A^2$Nav achieves promising ZS-VLN performance and even surpasses the supervised learning methods on R2R-Habitat and RxR-Habitat datasets.
Vision-and-language navigation (VLN) requires an embodied agent to navigate in realistic 3D environments using natural language instructions. Existing VLN methods suffer from training on small-scale environments or unreasonable path-instruction datasets, limiting the generalization to unseen environments. There are massive house tour videos on YouTube, providing abundant real navigation experiences and layout information. However, these videos have not been explored for VLN before. In this paper, we propose to learn an agent from these videos by creating a large-scale dataset which comprises reasonable path-instruction pairs from house tour videos and pre-training the agent on it. To achieve this, we have to tackle the challenges of automatically constructing path-instruction pairs and exploiting real layout knowledge from raw and unlabeled videos. To address these, we first leverage an entropy-based method to construct the nodes of a path trajectory. Then, we propose an action-aware generator for generating instructions from unlabeled trajectories. Last, we devise a trajectory judgment pretext task to encourage the agent to mine the layout knowledge. Experimental results show that our method achieves state-of-the-art performance on two popular benchmarks (R2R and REVERIE). Code is available at https://github.com/JeremyLinky/YouTube-VLN
We study the task of weakly-supervised point cloud semantic segmentation with sparse annotations (e.g., less than 0.1% points are labeled), aiming to reduce the expensive cost of dense annotations. Unfortunately, with extremely sparse annotated points, it is very difficult to extract both contextual and object information for scene understanding such as semantic segmentation. Motivated by masked modeling (e.g., MAE) in image and video representation learning, we seek to endow the power of masked modeling to learn contextual information from sparsely-annotated points. However, directly applying MAE to 3D point clouds with sparse annotations may fail to work. First, it is nontrivial to effectively mask out the informative visual context from 3D point clouds. Second, how to fully exploit the sparse annotations for context modeling remains an open question. In this paper, we propose a simple yet effective Contextual Point Cloud Modeling (CPCM) method that consists of two parts: a region-wise masking (RegionMask) strategy and a contextual masked training (CMT) method. Specifically, RegionMask masks the point cloud continuously in geometric space to construct a meaningful masked prediction task for subsequent context learning. CMT disentangles the learning of supervised segmentation and unsupervised masked context prediction for effectively learning the very limited labeled points and mass unlabeled points, respectively. Extensive experiments on the widely-tested ScanNet V2 and S3DIS benchmarks demonstrate the superiority of CPCM over the state-of-the-art.
Neural Radiance Fields (NeRF) is a revolutionary approach for rendering scenes by sampling a single ray per pixel and it has demonstrated impressive capabilities in novel-view synthesis from static scene images. However, in practice, we usually need to recover NeRF from unconstrained image collections, which poses two challenges: 1) the images often have dynamic changes in appearance because of different capturing time and camera settings; 2) the images may contain transient objects such as humans and cars, leading to occlusion and ghosting artifacts. Conventional approaches seek to address these challenges by locally utilizing a single ray to synthesize a color of a pixel. In contrast, humans typically perceive appearance and objects by globally utilizing information across multiple pixels. To mimic the perception process of humans, in this paper, we propose Cross-Ray NeRF (CR-NeRF) that leverages interactive information across multiple rays to synthesize occlusion-free novel views with the same appearances as the images. Specifically, to model varying appearances, we first propose to represent multiple rays with a novel cross-ray feature and then recover the appearance by fusing global statistics, i.e., feature covariance of the rays and the image appearance. Moreover, to avoid occlusion introduced by transient objects, we propose a transient objects handler and introduce a grid sampling strategy for masking out the transient objects. We theoretically find that leveraging correlation across multiple rays promotes capturing more global information. Moreover, extensive experimental results on large real-world datasets verify the effectiveness of CR-NeRF.
Adversarial detection aims to determine whether a given sample is an adversarial one based on the discrepancy between natural and adversarial distributions. Unfortunately, estimating or comparing two data distributions is extremely difficult, especially in high-dimension spaces. Recently, the gradient of log probability density (a.k.a., score) w.r.t. the sample is used as an alternative statistic to compute. However, we find that the score is sensitive in identifying adversarial samples due to insufficient information with one sample only. In this paper, we propose a new statistic called expected perturbation score (EPS), which is essentially the expected score of a sample after various perturbations. Specifically, to obtain adequate information regarding one sample, we perturb it by adding various noises to capture its multi-view observations. We theoretically prove that EPS is a proper statistic to compute the discrepancy between two samples under mild conditions. In practice, we can use a pre-trained diffusion model to estimate EPS for each sample. Last, we propose an EPS-based adversarial detection (EPS-AD) method, in which we develop EPS-based maximum mean discrepancy (MMD) as a metric to measure the discrepancy between the test sample and natural samples. We also prove that the EPS-based MMD between natural and adversarial samples is larger than that among natural samples. Extensive experiments show the superior adversarial detection performance of our EPS-AD.
Source-free Unsupervised Domain Adaptation (SF-UDA) aims to adapt a well-trained source model to an unlabeled target domain without access to the source data. One key challenge is the lack of source data during domain adaptation. To handle this, we propose to mine the hidden knowledge of the source model and exploit it to generate source avatar prototypes. To this end, we propose a Contrastive Prototype Generation and Adaptation (CPGA) method. CPGA consists of two stages: Prototype generation and Prototype adaptation. Extensive experiments on three UDA benchmark datasets demonstrate the superiority of CPGA. However, existing SF.UDA studies implicitly assume balanced class distributions for both the source and target domains, which hinders their real applications. To address this issue, we study a more practical SF-UDA task, termed imbalance-agnostic SF-UDA, where the class distributions of both the unseen source domain and unlabeled target domain are unknown and could be arbitrarily skewed. This task is much more challenging than vanilla SF-UDA due to the co-occurrence of covariate shifts and unidentified class distribution shifts between the source and target domains. To address this task, we extend CPGA and propose a new Target-aware Contrastive Prototype Generation and Adaptation (T-CPGA) method. Specifically, for better prototype adaptation in the imbalance-agnostic scenario, T-CPGA applies a new pseudo label generation strategy to identify unknown target class distribution and generate accurate pseudo labels, by utilizing the collective intelligence of the source model and an additional contrastive language-image pre-trained model. Meanwhile, we further devise a target label-distribution-aware classifier to adapt the model to the unknown target class distribution. We empirically show that T-CPGA significantly outperforms CPGA and other SF-UDA methods in imbalance-agnostic SF-UDA.