Shortcut learning is when a model -- e.g. a cardiac disease classifier -- exploits correlations between the target label and a spurious shortcut feature, e.g. a pacemaker, to predict the target label based on the shortcut rather than real discriminative features. This is common in medical imaging, where treatment and clinical annotations correlate with disease labels, making them easy shortcuts to predict disease. We propose a novel detection and quantification of the impact of potential shortcut features via diffusion-based counterfactual image generation that can synthetically remove or add shortcuts. Via a novel self-optimized masking scheme we spatially limit the changes made with no extra inference step, encouraging the removal of spatially constrained shortcut features while ensuring that the shortcut-free counterfactuals preserve their remaining image features to a high degree. Using these, we assess how shortcut features influence model predictions. This is enabled by our second contribution: An efficient diffusion-based counterfactual generation with significant inference speed-up at comparable counterfactual quality as state-of-the-art. We confirm our performance on two large chest X-ray datasets, a skin lesion dataset, and CelebA.
论文检索
输入标题、作者或关键词，从 2,387 篇学术成果中精准定位
Artistic images typically contain the unique creative styles of artists. However, it is easy to transfer an artist's style to arbitrary target images using style transfer techniques. To protect styles, some researchers use adversarial attacks to safeguard artists' artistic style images. Prior methods only considered defending against all style transfer models, but artists may allow specific models to transfer their artistic styles properly. To meet such requirements, we propose an Artistic Style Protection Scheme (ASPS). The scheme utilizes adversarial perturbations to introduce biases in the mean and variance of content and style features extracted by unauthorized models while aligning authorized models' content and style features. Additionally, it employs pixel-level and feature-level losses to enhance and degrade the output quality of authorized and unauthorized models, respectively. ASPS requires training only once; during usage, there is no need to see any style transfer models again. Meanwhile, it ensures that the visual quality of the authorized model is unaffected by perturbations. Experimental results demonstrate that our method effectively defends against unauthorized models' indiscriminate use of artistic styles, allowing authorized models to operate normally, thus effectively resolving the issue of controlled authorization regarding artists' artistic styles.
Videos are highly redundant data source and it is often enough to identify a few key moments to solve any given task. In this paper, we present a text-conditioned video resampler (TCR) module that uses a pre-trained and frozen visual encoder and large language model (LLM) to process long video sequences for a task. TCR localises relevant visual features from the video given a text condition and provides them to a LLM to generate a text response. Due to its lightweight design and use of cross-attention, TCR can process more than 100 frames at a time allowing the model to use much longer chunks of video than earlier works. We make the following contributions: (i) we design a transformer-based sampling architecture that can process long videos conditioned on a task, together with a training method that enables it to bridge pre-trained visual and language models; (ii) we empirically validate its efficacy on a wide variety of evaluation tasks, and set a new state-of-the-art on NextQA, EgoSchema, and the EGO4D-LTA challenge; and (iii) we determine tasks which require longer video contexts and that can thus be used effectively for further evaluation of long-range video models.
Learning Where to Look: Self-supervised Viewpoint Selection for Active Localization using Geometrical Information
PDF ↗Accurate localization in diverse environments is a fundamental challenge in computer vision and robotics. The task involves determining a sensor's precise position and orientation, typically a camera, within a given space. Traditional localization methods often rely on passive sensing, which may struggle in scenarios with limited features or dynamic environments. In response, this paper explores the domain of active localization, emphasizing the importance of viewpoint selection to enhance localization accuracy. Our contributions involve using a data-driven approach with a simple architecture designed for real-time operation, a self-supervised data training method, and the capability to consistently integrate our map into a planning framework tailored for real-world robotics applications. Our results demonstrate that our method performs better than the existing one target similar problems, generalizing on synthetic and real data. We also release an open-source implementation to benefit the community.
Missing Modality Prediction for Unpaired Multimodal Learning via Joint Embedding of Unimodal Models
PDF ↗Multimodal learning typically relies on the assumption that all modalities are fully available during both the training and inference phases. However, in real-world scenarios, consistently acquiring complete multimodal data presents significant challenges due to various factors. This often leads to the issue of missing modalities, where data for certain modalities are absent, posing considerable obstacles not only for the availability of multimodal pretrained models but also for their fine-tuning and the preservation of robustness in downstream tasks. To address these challenges, we propose a novel framework integrating parameter-efficient fine-tuning of unimodal pretrained models with a self-supervised joint-embedding learning method. This framework enables the model to predict the embedding of a missing modality in the representation space during inference. Our method effectively predicts the missing embedding through prompt tuning, leveraging information from available modalities and enhancing inter-modal interaction. We evaluate our approach on several multimodal benchmark datasets and demonstrate its effectiveness and robustness across various scenarios of missing modalities.
While conversational generative AI has shown considerable potential in enhancing decision-making for agricultural professionals, its exploration has predominantly been anchored in text-based interactions. The evolution of multimodal conversational AI, leveraging vast amounts of image-text data from diverse sources, marks a significant stride forward. However, the application of such advanced vision-language models in the agricultural domain, particularly for crop disease diagnosis, remains underexplored. In this work, we present the crop disease domain multimodal (CDDM) dataset, a pioneering resource designed to advance the field of agricultural research through the application of multimodal learning techniques. The dataset comprises 137,000 images of various crop diseases, accompanied by 1 million question-answer pairs that span a broad spectrum of agricultural knowledge, from disease identification to management practices. By integrating visual and textual data, CDDM facilitates the development of sophisticated question-answering systems capable of providing precise, useful advice to farmers and agricultural professionals. We demonstrate the utility of the dataset by finetuning state-of-the-art multimodal models, showcasing significant improvements in crop disease diagnosis. Specifically, we employed a novel finetuning strategy that utilizes low-rank adaptation (LoRA) to finetune the visual encoder, adapter and language model simultaneously. Our contributions include not only the dataset but also a finetuning strategy and a benchmark to stimulate further research in agricultural technology, aiming to bridge the gap between advanced AI techniques and practical agricultural applications.
CLIP, as a vision-language model, has significantly advanced Open-Vocabulary Semantic Segmentation (OVSS) with its zero-shot capabilities. Despite its success, its application to OVSS faces challenges due to its initial image-level alignment training, which affects its performance in tasks requiring detailed local context. Our study delves into the impact of CLIP's [CLS] token on patch feature correlations, revealing a dominance of "global" patches that hinders local feature discrimination. To overcome this, we propose CLIPtrase, a novel training-free semantic segmentation strategy that enhances local feature awareness through recalibrated self-correlation among patches. This approach demonstrates notable improvements in segmentation accuracy and the ability to maintain semantic coherence across objects. Experiments show that we are 22.3% ahead of CLIP on average on 9 segmentation benchmarks, outperforming existing state-of-the-art training-free methods.
The concept bottleneck model (CBM) is an interpretable-by-design framework that makes decisions by first predicting a set of interpretable concepts, and then predicting the class label based on the given concepts. Existing CBMs are trained with a fixed set of concepts (concepts are either annotated by the dataset or queried from language models). However, this closed-world assumption is unrealistic in practice, as users may wonder about the role of any desired concept in decision-making after the model is deployed. Inspired by the large success of recent vision-language pre-trained models such as CLIP in zero-shot classification, we propose OpenCBM'' to equip the CBM with open vocabulary concepts via: (1) Aligning the feature space of a trainable image feature extractor with that of a CLIP's image encoder via a prototype based feature alignment; (2) Simultaneously training an image classifier on the downstream dataset; (3) Reconstructing the trained classification head via any set of user-desired textual concepts encoded by CLIP's text encoder. To reveal potentially missing concepts from users, we further propose to iteratively find the closest concept embedding to the residual parameters during the reconstruction until the residual is small enough. To the best of our knowledge, ourOpenCBM'' is the first CBM with concepts of open vocabularies, providing users the unique benefit such as removing, adding, or replacing any desired concept to explain the model's prediction even after a model is trained. Moreover, our model significantly outperforms the previous state-of-the-art CBM by 9% in the classification accuracy on the benchmark dataset CUB-200-2011. The code will be released upon acceptance.
Fake It till You Make It: Curricular Dynamic Forgery Augmentations towards General Deepfake Detection
PDF ↗Previous studies in deepfake detection have shown promising results when testing face forgeries from the same dataset as the training. However, the problem remains challenging when one tries to generalize the detector to forgeries from unseen datasets and created by unseen methods. Utilizing forgery augmentation is one of the powerful lines to improve generalization performances. In this work, we present a novel general deepfake detection method, called Curricular Dynamic Forgery Augmentation (CDFA), which jointly trains a deepfake detector with a forgery augmentation policy network. Unlike the previous works, we propose to progressively apply forgery augmentations following a monotonic curriculum during the training. We further propose a dynamic forgery searching strategy to select one suitable forgery augmentation operation for each image varying between training stages, producing a forgery augmentation policy optimized for better generalization. In addition, we propose a novel forgery augmentation named self-shifted blending image to simply imitate the temporal inconsistency of deepfake generation. Comprehensive experiments show that CDFA can significantly improve both cross-datasets and cross-manipulations performances of various naive deepfake detectors in a plug-and-play way, and make them attain superior performances over the existing methods in several benchmark datasets. To facilitate the reproducible research, we will release our code upon the acceptance of the paper.
We introduce Adversarial Diffusion Distillation (ADD), a novel training approach that efficiently samples large-scale foundational image diffusion models in just 1--4 steps while maintaining high image quality. We use score distillation to leverage large-scale off-the-shelf image diffusion models as a teacher signal in combination with an adversarial loss to ensure high image fidelity even in the low-step regime of one or two sampling steps. Our analyses show that our model clearly outperforms existing few-step methods (GANs, Latent Consistency Models) in a single step and reaches the performance of state-of-the-art diffusion models (SDXL) in only four steps. ADD is the first method to unlock single-step, real-time image synthesis with foundation models.
Fast Adversarial Training (FAT) has gained increasing attention within the research community owing to its efficacy in improving adversarial robustness. Particularly noteworthy is the challenge posed by catastrophic overfitting (CO) in this field. Although existing FAT approaches have made strides in mitigating CO, the ascent of adversarial robustness occurs with a non-negligible decline in classification accuracy on clean samples. To tackle this issue, we initially employ the feature activation differences between clean and adversarial examples to analyze the underlying causes of CO. Intriguingly, our findings reveal that CO can be attributed to the feature coverage induced by a few specific pathways. By intentionally manipulating feature activation differences in these pathways with well-designed regularization terms, we can effectively mitigate and induce CO, providing further evidence for this observation. Notably, models trained stably with these terms exhibit superior performance compared to prior FAT work. On this basis, we harness CO to achieve `attack obfuscation', aiming to bolster model performance. Consequently, the models suffering from CO can attain optimal classification accuracy on both clean and adversarial data when adding random noise to inputs during evaluation. We also validate their robustness against transferred adversarial examples and the necessity of inducing CO to improve robustness. Hence, CO may not be a problem that has to be solved.
Enhancing Semantic Fidelity in Text-to-Image Synthesis: Attention Regulation in Diffusion Models
PDF ↗Recent advancements in diffusion models have notably improved the perceptual quality of generated images in text-to-image synthesis tasks. However, diffusion models often struggle to produce images that accurately reflect the intended semantics of the associated text prompts. We examine cross-attention layers in diffusion models and observe a propensity for these layers to disproportionately focus on certain tokens during the generation process, thereby undermining semantic fidelity. To address the issue of dominant attention, we introduce attention regulation, a computation-efficient on-the-fly optimization approach at inference time to align attention maps with the input text prompt. Notably, our method requires no additional training or fine-tuning and serves as a plug-in module on a model. Hence, the generation capacity of the original model is fully preserved. We compare our approach with alternative approaches across various datasets, evaluation metrics, and diffusion models. Experiment results show that our method consistently outperforms other baselines, yielding images that more faithfully reflect the desired concepts with reduced computation overhead.
We propose an application of online hard sample mining for efficient training of Neural Radiance Fields (NeRF). NeRF models produce state-of-the-art quality for many 3D reconstruction and rendering tasks but require substantial computational resources. The encoding of the scene information within the NeRF network parameters necessitates stochastic sampling. We observe that during the training, a major part of the compute time and memory usage is spent on processing already learnt samples, which no longer affect the model update significantly. We identify the backward pass on the stochastic samples as the computational bottleneck during the optimization. We thus perform the first forward pass in inference mode as a relatively low-cost search for hard samples. This is followed by building the computational graph and updating the NeRF network parameters using only the hard samples. To demonstrate the effectiveness of the proposed approach, we apply our method to Instant-NGP, resulting in significant improvements of the view-synthesis quality over the baseline (1 dB improvement on average per training time, or 2x speedup to reach the same PSNR level) along with 40% memory savings coming from using only the hard samples to build the computational graph. As our method only interfaces with the network module, we expect it to be widely applicable.
Assembling 3D objects from primitive bricks is challenging due to complex constraints and numerous possible combinations. Recent studies have demonstrated promising results on sequential LEGO brick assembly by graph modeling. However, existing approaches are class-specific and require significant computational and 3D annotation resources. In this work, we first propose a computationally efficient breadth-first search (BFS) LEGO-Tree structure to model sequential assembly actions. Based on the LEGO-Tree structure, we then design a class-agnostic tree-transformer framework to predict assembly actions from multi-view images. A major challenge is the costly acquisition of step-wise action labels. We address this by leveraging synthetic-to-real transfer learning. Specifically, our model pre-trains on synthetic data with full action label supervision. We further circumvent the requirement for real data action labels by introducing an action-to-silhouette projection for self-supervision. With no real data annotation, our model outperforms existing methods with 3D supervision by 7.8% and 11.3% in mIoU on the MNIST and ModelNet Construction datasets, respectively.
Radar-based perception has gained increasing attention in autonomous driving, yet the inherent sparsity of radars poses challenges. Radar raw data often contains excessive noise, whereas radar point clouds retain only limited information. In this work, we holistically treat the sparse nature of radar data by introducing an adaptive subsampling method together with a tailored network architecture that exploits the sparsity patterns to discover global and local dependencies in the radar signal. Our subsampling module selects a subset of pixels from range-doppler (RD) spectra that contribute most to the downstream perception tasks. To improve the feature extraction on sparse subsampled data, we propose a new way of applying graph neural networks on radar data and design a novel two-branch backbone to capture both global and local neighbor information. An attentive fusion module is applied to combine features from both branches. Experiments on the RADIal dataset show that our SparseRadNet exceeds state-of-the-art (SOTA) performance in object detection and achieves close to SOTA accuracy in freespace segmentation, meanwhile using sparse subsampled input data.
Text-driven 3D texturing requires the generation of high-fidelity texture that conforms to given geometry and description. Recently, the high-quality text-to-image generation ability of 2D diffusion model has significantly promoted this task, by converting it into a texture optimization process guided by multi-view synthesized images, where the generation of high-quality and multi-view consistency images becomes the key issue. State-of-the-art methods achieve the consistency between different views by treating image generation on a novel view as image inpainting conditioned on the texture generated by previously views. However, due to the accumulated semantic divergence of local inpainting and the occlusion between object parts on sparse views, these inpainting-based methods often fail to deal with long-range texture consistency. To address these, we present P3G, a texturing approach based on learned Pseudo 3D Guidance. The key idea of P3G is to first learn a coarse but consistent texture, to serve as a global semantics guidance for encouraging the consistency between images generated on different views. To this end, we incorporate pre-trained text-to-image diffusion models and multi-view optimization to achieve propagating accurate semantics globally for leaning the guidance, and design an efficient framework for high-quality and multi-view consistent image generation that integrates the learned semantic guidance. Quantitative and qualitative evaluation on variant 3D shapes demonstrates the superiority of our P3G on both consistency and overall visual quality.
We introduce Spline-based Transformers, a novel class of Transformer models that eliminate the need for positional encoding. Inspired by workflows using splines in computer animation, our Spline-based Transformer embeds an input sequence of elements as a smooth trajectory in latent space. Overcoming drawbacks of positional encoding such as sequence length extrapolation, Spline-based Transformers also provide a novel way for users to interact with transformer latent spaces by directly manipulating the latent control points to create new latent trajectories and sequences. We demonstrate the superior performance of our approach in comparison to conventional positional encoding on a variety of datasets, ranging from synthetic 2D to large-scale real-world datasets of images, 3D shapes, and animations.
We revisit the problem of camera relocalization under memory budget through a combination of product quantization and map compression. This achieves high compression rates but leads to performance drop. To address memory performance tradeoff, we train a light-weight scene-specific auto-encoder network that performs quantization-dequantization in an end-to-end differentiable manner updating both product quantization centroids and network parameters. Unlike standard L2 reconstruction loss for training auto-encoder network, we show that additional margin-based metric losses are key to achieve good performance. Results show that for a descriptor memory of 1 MB, we can achieve competitive performance on Aachen Day with only 5 % drop in performance.
We delve into a crucial yet often overlooked challenge inherent to Vision Transformers (ViTs): feature maps of these models exhibit grid-like artifacts, which hurt the performance of ViTs in downstream dense prediction tasks such as segmentation, depth prediction, and object discovery. We trace this fundamental issue down to the positional embeddings at the input stage. we propose a two-stage denoising approach, termed Denoising Vision Transformers (DVT). In the first stage, we separate the clean features from those contaminated by positional artifacts by enforcing cross-view feature consistency with neural fields on a per-image basis. This per-image optimization process extracts artifact-free features from raw ViT outputs, providing clean feature estimates for offline applications. In the second stage, we train a lightweight Transformer block to predict clean features from raw ViT outputs, leveraging the derived estimates of the clean features as supervision. Our DVT does not require re-training the existing pre-trained ViTs, and is immediately applicable to any Vision Transformer architecture. We evaluate our method on a variety of representative ViTs (DINO, DeiT-III, EVA02, CLIP, DINOv2, DINOv2-reg) and demonstrate that our DVT consistently and significantly improves existing state-of-the-art general-purpose models in semantic and geometric tasks across multiple datasets. We hope our study will encourage a re-evaluation of ViT design, especially regarding the naive use of positional embeddings. Our code and models will be released.
Nowadays an ever-growing concerning phenomenon, the emergence of algorithmic biases that can lead to unfair models, emerges. Several debiasing approaches have been proposed in the realm of deep learning, employing more or less sophisticated approaches to discourage these models from massively employing these biases. However, a question emerges: is this extra complexity really necessary? Is a vanilla-trained model already embodying some ``unbiased sub-networks'' that can be used in isolation and propose a solution without relying on the algorithmic biases? In this work, we show that such a sub-network typically exists, and can be extracted from a vanilla-trained model without requiring additional training. We further validate that such specific architecture is incapable of learning a specific bias, suggesting that there are possible architectural countermeasures to the problem of biases in deep neural networks.