Untrained neural networks pioneered by Deep Image Prior (DIP) have recently enabled MRI reconstruction without requiring fully-sampled measurements for training. Their success is widely attributed to the implicit regularization induced by suitable network architectures. However, the lack of understanding of such architectural priors results in superfluous design choices and sub-optimal outcomes. This work aims to simplify the architectural design decisions for DIP-MRI to facilitate its practical deployment. We observe that certain architectural components are more prone to causing overfitting regardless of the number of parameters, incurring severe reconstruction artifacts by hindering accurate extrapolation on the un-acquired measurements. We interpret this phenomenon from a frequency perspective and find that the architectural characteristics favoring low frequencies, i.e., deep and narrow with unlearnt upsampling, can lead to enhanced generalization and hence better reconstruction. Building on this insight, we propose two architecture-agnostic remedies: one to constrain the frequency range of the white-noise input and the other to penalize the Lipschitz constants of the network. We demonstrate that even with just one extra line of code on the input, the performance gap between the ill-designed models and the high-performing ones can be closed. These results signify that for the first time, architectural biases on untrained MRI reconstruction can be mitigated without architectural modifications.
Facial animation in virtual reality environments is essential for applications that necessitate clear visibility of the user's face and the ability to convey emotional signals. In our scenario, we animate the face of an operator who controls a robotic Avatar system. The use of facial animation is particularly valuable when the perception of interacting with a specific individual, rather than just a robot, is intended. Purely keypoint-driven animation approaches struggle with the complexity of facial movements. We present a hybrid method that uses both keypoints and direct visual guidance from a mouth camera. Our method generalizes to unseen operators and requires only a quick enrolment step with capture of two short videos. Multiple source images are selected with the intention to cover different facial expressions. Given a mouth camera frame from the HMD, we dynamically construct the target keypoints and apply an attention mechanism to determine the importance of each source image. To resolve keypoint ambiguities and animate a broader range of mouth expressions, we propose to inject visual mouth camera information into the latent space. We enable training on large-scale speaking head datasets by simulating the mouth camera input with its perspective differences and facial deformations. Our method outperforms a baseline in quality, capability, and temporal consistency. In addition, we highlight how the facial animation contributed to our victory at the ANA Avatar XPRIZE Finals.
Large-scale language-image pre-trained models (e.g., CLIP) have shown superior performances on many cross-modal retrieval tasks. However, the problem of transferring the knowledge learned from such models to video-based person re-identification (ReID) has barely been explored. In addition, there is a lack of decent text descriptions in current ReID benchmarks. To address these issues, in this work, we propose a novel one-stage text-free CLIP-based learning framework named TF-CLIP for video-based person ReID. More specifically, we extract the identity-specific sequence feature as the CLIP-Memory to replace the text feature. Meanwhile, we design a Sequence-Specific Prompt (SSP) module to update the CLIP-Memory online. To capture temporal information, we further propose a Temporal Memory Diffusion (TMD) module, which consists of two key components: Temporal Memory Construction (TMC) and Memory Diffusion (MD). Technically, TMC allows the frame-level memories in a sequence to communicate with each other, and to extract temporal information based on the relations within the sequence. MD further diffuses the temporal memories to each token in the original features to obtain more robust sequence features. Extensive experiments demonstrate that our proposed method shows much better results than other state-of-the-art methods on MARS, LS-VID and iLIDS-VID. The code is available at https://github.com/AsuradaYuci/TF-CLIP.
Multi-modal aspect-based sentiment analysis (MABSA) has recently attracted increasing attention. The span-based extraction methods, such as FSUIE, demonstrate strong performance in sentiment analysis due to their joint modeling of input sequences and target labels. However, previous methods still have certain limitations: (i) They ignore the difference in the focus of visual information between different analysis targets (aspect or sentiment). (ii) Combining features from uni-modal encoders directly may not be sufficient to eliminate the modal gap and can cause difficulties in capturing the image-text pairwise relevance. (iii) Existing span-based methods for MABSA ignore the pairwise relevance of target span boundaries. To tackle these limitations, we propose a novel framework called DQPSA for multi-modal sentiment analysis. Specifically, our model contains a Prompt as Dual Query (PDQ) module that uses the prompt as both a visual query and a language query to extract prompt-aware visual information and strengthen the pairwise relevance between visual information and the analysis target. Additionally, we introduce an Energy-based Pairwise Expert (EPE) module that models the boundaries pairing of the analysis target from the perspective of an Energy-based Model. This expert predicts aspect or sentiment span based on pairwise stability. Experiments on three widely used benchmarks demonstrate that DQPSA outperforms previous approaches and achieves a new state-of-the-art performance.
Implicit neural representation has demonstrated promising results in view synthesis for large and complex scenes. However, existing approaches either fail to capture the fast-moving objects or need to build the scene graph without camera ego-motions, leading to low-quality synthesized views of the scene. We aim to jointly solve the view synthesis problem of large-scale urban scenes and fast-moving vehicles, which is more practical and challenging. To this end, we first leverage a graph structure to learn the local scene representations of dynamic objects and the background. Then, we design a progressive scheme that dynamically allocates a new local scene graph trained with frames within a temporal window, allowing us to scale up the representation to an arbitrarily large scene. Besides, the training views of urban scenes are relatively sparse, which leads to a significant decline in reconstruction accuracy for dynamic objects. Therefore, we design a frequency auto-encoder network to encode the latent code and regularize the frequency range of objects, which can enhance the representation of dynamic objects and address the issue of sparse image inputs. Additionally, we employ lidar point projection to maintain geometry consistency in large-scale urban scenes. Experimental results demonstrate that our method achieves state-of-the-art view synthesis accuracy, object manipulation, and scene roaming ability. The code will be open-sourced upon paper acceptance.
Recent advancements in 3D reconstruction from single images have been driven by the evolution of generative models. Prominent among these are methods based on Score Distillation Sampling (SDS) and the adaptation of diffusion models in the 3D domain. Despite their progress, these techniques often face limitations due to slow optimization or rendering processes, leading to extensive training and optimization times. In this paper, we introduce a novel approach for single-view reconstruction that efficiently generates a 3D model from a single image via feed-forward inference. Our method utilizes two transformer-based networks, namely a point decoder and a triplane decoder, to reconstruct 3D objects using a hybrid Triplane-Gaussian intermediate representation. This hybrid representation strikes a balance, achieving a faster rendering speed compared to implicit representations while simultaneously delivering superior rendering quality than explicit representations. The point decoder is designed for generating point clouds from single images, offering an explicit representation which is then utilized by the triplane decoder to query Gaussian features for each point. This design choice addresses the challenges associated with directly regressing explicit 3D Gaussian attributes characterized by their non-structural nature. Subsequently, the 3D Gaussians are decoded by an MLP to enable rapid rendering through splatting. Both decoders are built upon a scalable, transformer-based architecture and have been efficiently trained on large-scale 3D datasets. The evaluations conducted on both synthetic datasets and real-world images demonstrate that our method not only achieves higher quality but also ensures a faster runtime in comparison to previous state-of-the-art techniques. Please see our project page at https://zouzx.github.io/TriplaneGaussian/.
Our work focuses on tackling large-scale fine-grained image retrieval as ranking the images depicting the concept of interests (i.e., the same sub-category labels) highest based on the fine-grained details in the query. It is desirable to alleviate the challenges of both fine-grained nature of small inter-class variations with large intra-class variations and explosive growth of fine-grained data for such a practical task. In this paper, we propose attribute-aware hashing networks with self-consistency for generating attribute-aware hash codes to not only make the retrieval process efficient, but also establish explicit correspondences between hash codes and visual attributes. Specifically, based on the captured visual representations by attention, we develop an encoder-decoder structure network of a reconstruction task to unsupervisedly distill high-level attribute-specific vectors from the appearance-specific visual representations without attribute annotations. Our models are also equipped with a feature decorrelation constraint upon these attribute vectors to strengthen their representative abilities. Then, driven by preserving original entities' similarity, the required hash codes can be generated from these attribute-specific vectors and thus become attribute-aware. Furthermore, to combat simplicity bias in deep hashing, we consider the model design from the perspective of the self-consistency principle and propose to further enhance models' self-consistency by equipping an additional image reconstruction path. Comprehensive quantitative experiments under diverse empirical settings on six fine-grained retrieval datasets and two generic retrieval datasets show the superiority of our models over competing methods.
Histopathology image analysis is the golden standard of clinical diagnosis for Cancers. In doctors daily routine and computer-aided diagnosis, the Whole Slide Image (WSI) of histopathology tissue is used for analysis. Because of the extremely large scale of resolution, previous methods generally divide the WSI into a large number of patches, then aggregate all patches within a WSI by Multi-Instance Learning (MIL) to make the slide-level prediction when developing computer-aided diagnosis tools. However, most previous WSI-MIL models using global-attention without pairwise interaction and any positional information, or self-attention with absolute position embedding can not well handle shape varying large WSIs, e.g. testing WSIs after model deployment may be larger than training WSIs, since the model development set is always limited due to the difficulty of histopathology WSIs collection. To deal with the problem, in this paper, we propose to amend position embedding for shape varying long-contextual WSI by introducing Linear Bias into Attention, and adapt it from 1-d long sequence into 2-d long-contextual WSI which helps model extrapolate position embedding to unseen or under-fitted positions. We further utilize Flash-Attention module to tackle the computational complexity of Transformer, which also keep full self-attention performance compared to previous attention approximation work. Our method, Long-contextual MIL (Long-MIL) are evaluated on extensive experiments including 4 dataset including WSI classification and survival prediction tasks to validate the superiority on shape varying WSIs. The source code will be open-accessed soon.
We propose a novel feed-forward 3D editing framework called Shap-Editor. Prior research on editing 3D objects primarily concentrated on editing individual objects by leveraging off-the-shelf 2D image editing networks. This is achieved via a process called distillation, which transfers knowledge from the 2D network to 3D assets. Distillation necessitates at least tens of minutes per asset to attain satisfactory editing results, and is thus not very practical. In contrast, we ask whether 3D editing can be carried out directly by a feed-forward network, eschewing test-time optimisation. In particular, we hypothesise that editing can be greatly simplified by first encoding 3D objects in a suitable latent space. We validate this hypothesis by building upon the latent space of Shap-E. We demonstrate that direct 3D editing in this space is possible and efficient by building a feed-forward editor network that only requires approximately one second per edit. Our experiments show that Shap-Editor generalises well to both in-distribution and out-of-distribution 3D assets with different prompts, exhibiting comparable performance with methods that carry out test-time optimisation for each edited instance.
Large language models have achieved great success in recent years, so as their variants in vision. Existing vision-language models can describe images in natural languages, answer visual-related questions, or perform complex reasoning about the image. However, it is yet unclear how localization tasks, such as word grounding or referring localization, can be performed using large language models. In this work, we aim to develop a vision-language model that can take locations, for example, a set of points or boxes, as either inputs or outputs. When taking locations as inputs, the model performs location-conditioned captioning, which generates captions for the indicated object or region. When generating locations as outputs, our model regresses pixel coordinates for each output word generated by the language model, and thus performs dense word grounding. Our model is pre-trained on the Localized Narrative dataset, which contains pixel-word-aligned captioning from human attention. We show our model can be applied to various location-aware vision-language tasks, including referring localization, location-conditioned captioning, and dense object captioning, archiving state-of-the-art performance on RefCOCO and Visual Genome. Project page: https://jerryxu.net/PixelLLM .