Text-to-image synthesis, a subfield of multimodal generation, has gained significant attention in recent years. We propose a cost-effective approach for image-to-prompt generation that leverages generative models to generate textual prompts without the need for large amounts of annotated data. We divide our method into two stages: online stage and offline stage. We use a combination of the CLIP model and K-nearest neighbors (KNN) algorithm. The proposed system consists of two main parts: an offline task and an online task. Our method owns the highest metric 0.612 among these models, which is 0.013, 0.055, 0.011 higher than Clip, Clip + KNN(top 10) respectively.
We propose to improve transformers of a specific modality with irrelevant data from other modalities, e.g., improve an ImageNet model with audio or point cloud datasets. We would like to highlight that the data samples of the target modality are irrelevant to the other modalities, which distinguishes our method from other works utilizing paired (e.g., CLIP) or interleaved data of different modalities. We propose a methodology named Multimodal Pathway - given a target modality and a transformer designed for it, we use an auxiliary transformer trained with data of another modality and construct pathways to connect components of the two models so that data of the target modality can be processed by both models. In this way, we utilize the universal sequence-to-sequence modeling abilities of transformers obtained from two modalities. As a concrete implementation, we use a modality-specific tokenizer and task-specific head as usual but utilize the transformer blocks of the auxiliary model via a proposed method named Cross-Modal Re-parameterization, which exploits the auxiliary weights without any inference costs. On the image, point cloud, video, and audio recognition tasks, we observe significant and consistent performance improvements with irrelevant data from other modalities. The code and models are available at https://github.com/AILab-CVC/M2PT.
Head-mounted displays (HMDs) serve as indispensable devices for observing extended reality (XR) environments and virtual content. However, HMDs present an obstacle to external recording techniques as they block the upper face of the user. This limitation significantly affects social XR applications, specifically teleconferencing, where facial features and eye gaze information play a vital role in creating an immersive user experience. In this study, we propose a new network for expression-aware video inpainting for HMD removal (EVI-HRnet) based on generative adversarial networks (GANs). Our model effectively fills in missing information with regard to facial landmarks and a single occlusion-free reference image of the user. The framework and its components ensure the preservation of the user's identity across frames using the reference frame. To further improve the level of realism of the inpainted output, we introduce a novel facial expression recognition (FER) loss function for emotion preservation. Our results demonstrate the remarkable capability of the proposed framework to remove HMDs from facial videos while maintaining the subject's facial expression and identity. Moreover, the outputs exhibit temporal consistency along the inpainted frames. This lightweight framework presents a practical approach for HMD occlusion removal, with the potential to enhance various collaborative XR applications without the need for additional hardware.
We propose a novel optimization-based human mesh recovery method from a single image. Given a test exemplar, previous approaches optimize the pre-trained regression network to minimize the 2D re-projection loss, which however suffer from over-/under-fitting problems. This is because the ``exemplar optimization'' at testing time has too weak relation to the pre-training process, and the exemplar optimization loss function is different from the training loss function. (1) We incorporate exemplar optimization into the training stage. During training, our method first executes exemplar optimization and subsequently proceeds with training-time optimization. The exemplar optimization may run into a wrong direction, while the subsequent training optimization serves to correct the deviation. Involved in training, the exemplar optimization learns to adapt its behavior to training data, thereby acquires generalibility to test exemplars. (2) We devise a dual-network architecture to convey the novel training paradigm, which is composed of a main regression network and an auxiliary network, in which we can formulate the exemplar optimization loss function in the same form as the training loss function. This further enhances the compatibility between the exemplar and training optimizations. Experiments demonstrate that our exemplar optimization after the novel training scheme significantly outperforms state-of-the-art approaches.
In High-definition (HD) maps, lane elements constitute the majority of components and demand stringent localization requirements to ensure safe vehicle navigation. Vision lane detection with LiDAR position assignment is a prevalent method to acquire initial lanes for HD maps. However, due to incorrect vision detection and coarse camera-LiDAR calibration, initial lanes may deviate from their true positions within an uncertain range. To mitigate the need for manual lane correction, we propose a patch-wise lane correction network (PLCNet) to automatically correct the positions of initial lane points in local LiDAR images that are transformed from point clouds. PLCNet first extracts multi-scale image features and crops patch (ROI) features centered at each initial lane point. By applying ROIAlign, the fix-sized ROI features are flattened into 1D features. Then, a 1D lane attention module is devised to compute instance-level lane features with adaptive weights. Finally, lane correction offsets are inferred by a multi-layer perceptron and used to correct the initial lane positions. Considering practical applications, our automatic method supports merging local corrected lanes into global corrected lanes. Through extensive experiments on a self-built dataset, we demonstrate that PLCNet achieves fast and effective initial lane correction.
Face recognition systems are widely deployed in high-security applications such as for biometric verification at border controls. Despite their high accuracy on pristine data, it is well-known that digital manipulations, such as face morphing, pose a security threat to face recognition systems. Malicious actors can exploit the facilities offered by the identity document issuance process to obtain identity documents containing morphed images. Thus, subjects who contributed to the creation of the morphed image can with high probability use the identity document to bypass automated face recognition systems. In recent years, no-reference (i.e., single image) and differential morphing attack detectors have been proposed to tackle this risk. These systems are typically evaluated in isolation from the face recognition system that they have to operate jointly with and do not consider the face recognition process. Contrary to most existing works, we present a novel method for adapting deep learning-based face recognition systems to be more robust against face morphing attacks. To this end, we introduce TetraLoss, a novel loss function that learns to separate morphed face images from its contributing subjects in the embedding space while still preserving high biometric verification performance. In a comprehensive evaluation, we show that the proposed method can significantly enhance the original system while also significantly outperforming other tested baseline methods.
The fusion of vision and language has brought about a transformative shift in computer vision through the emergence of Vision-Language Models (VLMs). However, the resource-intensive nature of existing VLMs poses a significant challenge. We need an accessible method for developing the next generation of VLMs. To address this issue, we propose Zoom-shot, a novel method for transferring the zero-shot capabilities of CLIP to any pre-trained vision encoder. We do this by exploiting the multimodal information (i.e. text and image) present in the CLIP latent space through the use of specifically designed multimodal loss functions. These loss functions are (1) cycle-consistency loss and (2) our novel prompt-guided knowledge distillation loss (PG-KD). PG-KD combines the concept of knowledge distillation with CLIP's zero-shot classification, to capture the interactions between text and image features. With our multimodal losses, we train a $\textbf{linear mapping}$ between the CLIP latent space and the latent space of a pre-trained vision encoder, for only a $\textbf{single epoch}$. Furthermore, Zoom-shot is entirely unsupervised and is trained using $\textbf{unpaired}$ data. We test the zero-shot capabilities of a range of vision encoders augmented as new VLMs, on coarse and fine-grained classification datasets, outperforming the previous state-of-the-art in this problem domain. In our ablations, we find Zoom-shot allows for a trade-off between data and compute during training; and our state-of-the-art results can be obtained by reducing training from 20% to 1% of the ImageNet training data with 20 epochs. All code and models are available on GitHub.
We consider the problem of source-free unsupervised category-level pose estimation from only RGB images to a target domain without any access to source domain data or 3D annotations during adaptation. Collecting and annotating real-world 3D data and corresponding images is laborious, expensive, yet unavoidable process, since even 3D pose domain adaptation methods require 3D data in the target domain. We introduce 3DUDA, a method capable of adapting to a nuisance-ridden target domain without 3D or depth data. Our key insight stems from the observation that specific object subparts remain stable across out-of-domain (OOD) scenarios, enabling strategic utilization of these invariant subcomponents for effective model updates. We represent object categories as simple cuboid meshes, and harness a generative model of neural feature activations modeled at each mesh vertex learnt using differential rendering. We focus on individual locally robust mesh vertex features and iteratively update them based on their proximity to corresponding features in the target domain even when the global pose is not correct. Our model is then trained in an EM fashion, alternating between updating the vertex features and the feature extractor. We show that our method simulates fine-tuning on a global pseudo-labeled dataset under mild assumptions, which converges to the target domain asymptotically. Through extensive empirical validation, including a complex extreme UDA setup which combines real nuisances, synthetic noise, and occlusion, we demonstrate the potency of our simple approach in addressing the domain shift challenge and significantly improving pose estimation accuracy.
Current large-scale diffusion models represent a giant leap forward in conditional image synthesis, capable of interpreting diverse cues like text, human poses, and edges. However, their reliance on substantial computational resources and extensive data collection remains a bottleneck. On the other hand, the integration of existing diffusion models, each specialized for different controls and operating in unique latent spaces, poses a challenge due to incompatible image resolutions and latent space embedding structures, hindering their joint use. Addressing these constraints, we present "PanGu-Draw", a novel latent diffusion model designed for resource-efficient text-to-image synthesis that adeptly accommodates multiple control signals. We first propose a resource-efficient Time-Decoupling Training Strategy, which splits the monolithic text-to-image model into structure and texture generators. Each generator is trained using a regimen that maximizes data utilization and computational efficiency, cutting data preparation by 48% and reducing training resources by 51%. Secondly, we introduce "Coop-Diffusion", an algorithm that enables the cooperative use of various pre-trained diffusion models with different latent spaces and predefined resolutions within a unified denoising process. This allows for multi-control image synthesis at arbitrary resolutions without the necessity for additional data or retraining. Empirical validations of Pangu-Draw show its exceptional prowess in text-to-image and multi-control image generation, suggesting a promising direction for future model training efficiencies and generation versatility. The largest 5B T2I PanGu-Draw model is released on the Ascend platform. Project page: $\href{https://pangu-draw.github.io}{this~https~URL}$
Recently, interpretable machine learning has re-explored concept bottleneck models (CBM), comprising step-by-step prediction of the high-level concepts from the raw features and the target variable from the predicted concepts. A compelling advantage of this model class is the user's ability to intervene on the predicted concept values, affecting the model's downstream output. In this work, we introduce a method to perform such concept-based interventions on already-trained neural networks, which are not interpretable by design, given an annotated validation set. Furthermore, we formalise the model's intervenability as a measure of the effectiveness of concept-based interventions and leverage this definition to fine-tune black-box models. Empirically, we explore the intervenability of black-box classifiers on synthetic tabular and natural image benchmarks. We demonstrate that fine-tuning improves intervention effectiveness and often yields better-calibrated predictions. To showcase the practical utility of the proposed techniques, we apply them to deep chest X-ray classifiers and show that fine-tuned black boxes can be as intervenable and more performant than CBMs.