In this work, we tackle the challenging problem of long-tailed image recognition. Previous long-tailed recognition approaches mainly focus on data augmentation or re-balancing strategies for the tail classes to give them more attention during model training. However, these methods are limited by the small number of training images for the tail classes, which results in poor feature representations. To address this issue, we propose the Latent Categories based long-tail Recognition (LCReg) method. Our hypothesis is that common latent features shared by head and tail classes can be used to improve feature representation. Specifically, we learn a set of class-agnostic latent features shared by both head and tail classes, and then use semantic data augmentation on the latent features to implicitly increase the diversity of the training sample. We conduct extensive experiments on five long-tailed image recognition datasets, and the results show that our proposed method significantly improves the baselines.
Recent neural networks based surface reconstruction can be roughly divided into two categories, one warping templates explicitly and the other representing 3D surfaces implicitly. To enjoy the advantages of both, we propose a novel 3D representation, Neural Vector Fields (NVF), which adopts the explicit learning process to manipulate meshes and implicit unsigned distance function (UDF) representation to break the barriers in resolution and topology. This is achieved by directly predicting the displacements from surface queries and modeling shapes as Vector Fields, rather than relying on network differentiation to obtain direction fields as most existing UDF-based methods do. In this way, our approach is capable of encoding both the distance and the direction fields so that the calculation of direction fields is differentiation-free, circumventing the non-trivial surface extraction step. Furthermore, building upon NVFs, we propose to incorporate two types of shape codebooks, \ie, NVFs (Lite or Ultra), to promote cross-category reconstruction through encoding cross-object priors. Moreover, we propose a new regularization based on analyzing the zero-curl property of NVFs, and implement this through the fully differentiable framework of our NVF (ultra). We evaluate both NVFs on four surface reconstruction scenarios, including watertight vs non-watertight shapes, category-agnostic reconstruction vs category-unseen reconstruction, category-specific, and cross-domain reconstruction.
Deep learning has proved to be very effective in video action recognition. Video violence recognition attempts to learn the human multi-dynamic behaviours in more complex scenarios. In this work, we develop a method for video violence recognition from a new perspective of skeleton points. Unlike the previous works, we first formulate 3D skeleton point clouds from human skeleton sequences extracted from videos and then perform interaction learning on these 3D skeleton point clouds. Specifically, we propose two types of Skeleton Points Interaction Learning (SPIL) strategies: (i) Local-SPIL: by constructing a specific weight distribution strategy between local regional points, Local-SPIL aims to selectively focus on the most relevant parts of them based on their features and spatial-temporal position information. In order to capture diverse types of relation information, a multi-head mechanism is designed to aggregate different features from independent heads to jointly handle different types of relationships between points. (ii) Global-SPIL: to better learn and refine the features of the unordered and unstructured skeleton points, Global-SPIL employs the self-attention layer that operates directly on the sampled points, which can help to make the output more permutation-invariant and well-suited for our task. Extensive experimental results validate the effectiveness of our approach and show that our model outperforms the existing networks and achieves new state-of-the-art performance on video violence datasets.
Recent strides in Text-to-3D techniques have been propelled by distilling knowledge from powerful large text-to-image diffusion models (LDMs). Nonetheless, existing Text-to-3D approaches often grapple with challenges such as over-saturation, inadequate detailing, and unrealistic outputs. This study presents a novel strategy that leverages explicitly synthesized multi-view images to address these issues. Our approach involves the utilization of image-to-image pipelines, empowered by LDMs, to generate posed high-quality images based on the renderings of coarse 3D models. Although the generated images mostly alleviate the aforementioned issues, challenges such as view inconsistency and significant content variance persist due to the inherent generative nature of large diffusion models, posing extensive difficulties in leveraging these images effectively. To overcome this hurdle, we advocate integrating a discriminator alongside a novel Diffusion-GAN dual training strategy to guide the training of 3D models. For the incorporated discriminator, the synthesized multi-view images are considered real data, while the renderings of the optimized 3D models function as fake data. We conduct a comprehensive set of experiments that demonstrate the effectiveness of our method over baseline approaches.
We study the problem of synthesizing a long-term dynamic video from only a single image. This is challenging since it requires consistent visual content movements given large camera motions. Existing methods either hallucinate inconsistent perpetual views or struggle with long camera trajectories. To address these issues, it is essential to estimate the underlying 4D (including 3D geometry and scene motion) and fill in the occluded regions. To this end, we present Make-It-4D, a novel method that can generate a consistent long-term dynamic video from a single image. On the one hand, we utilize layered depth images (LDIs) to represent a scene, and they are then unprojected to form a feature point cloud. To animate the visual content, the feature point cloud is displaced based on the scene flow derived from motion estimation and the corresponding camera pose. Such 4D representation enables our method to maintain the global consistency of the generated dynamic video. On the other hand, we fill in the occluded regions by using a pretrained diffusion model to inpaint and outpaint the input image. This enables our method to work under large camera motions. Benefiting from our design, our method can be training-free which saves a significant amount of training time. Experimental results demonstrate the effectiveness of our approach, which showcases compelling rendering results.
The remarkable multimodal capabilities demonstrated by OpenAI's GPT-4 have sparked significant interest in the development of multimodal Large Language Models (LLMs). A primary research objective of such models is to align visual and textual modalities effectively while comprehending human instructions. Current methodologies often rely on annotations derived from benchmark datasets to construct image-dialogue datasets for training purposes, akin to instruction tuning in LLMs. However, these datasets often exhibit domain bias, potentially constraining the generative capabilities of the models. In an effort to mitigate these limitations, we propose a novel data collection methodology that synchronously synthesizes images and dialogues for visual instruction tuning. This approach harnesses the power of generative models, marrying the abilities of ChatGPT and text-to-image generative models to yield a diverse and controllable dataset with varied image content. This not only provides greater flexibility compared to existing methodologies but also significantly enhances several model capabilities. Our research includes comprehensive experiments conducted on various datasets using the open-source LLAVA model as a testbed for our proposed pipeline. Our results underscore marked enhancements across more than ten commonly assessed capabilities,
Dark videos often lose essential information, which causes the knowledge learned by networks is not enough to accurately recognize actions. Existing knowledge assembling methods require massive GPU memory to distill the knowledge from multiple teacher models into a student model. In action recognition, this drawback becomes serious due to much computation required by video process. Constrained by limited computation source, these approaches are infeasible. To address this issue, we propose an unlimited knowledge distillation (UKD) in this paper. Compared with existing knowledge assembling methods, our UKD can effectively assemble different knowledge without introducing high GPU memory consumption. Thus, the number of teaching models for distillation is unlimited. With our UKD, the network's learned knowledge can be remarkably enriched. Our experiments show that the single stream network distilled with our UKD even surpasses a two-stream network. Extensive experiments are conducted on the ARID dataset.
The key to the success of few-shot segmentation (FSS) lies in how to effectively utilize support samples. Most solutions compress support foreground (FG) features into prototypes, but lose some spatial details. Instead, others use cross attention to fuse query features with uncompressed support FG. Query FG could be fused with support FG, however, query background (BG) cannot find matched BG features in support FG, yet inevitably integrates dissimilar features. Besides, as both query FG and BG are combined with support FG, they get entangled, thereby leading to ineffective segmentation. To cope with these issues, we design a self-calibrated cross attention (SCCA) block. For efficient patch-based attention, query and support features are firstly split into patches. Then, we design a patch alignment module to align each query patch with its most similar support patch for better cross attention. Specifically, SCCA takes a query patch as Q, and groups the patches from the same query image and the aligned patches from the support image as K&V. In this way, the query BG features are fused with matched BG features (from query patches), and thus the aforementioned issues will be mitigated. Moreover, when calculating SCCA, we design a scaled-cosine mechanism to better utilize the support features for similarity calculation. Extensive experiments conducted on PASCAL-5^i and COCO-20^i demonstrate the superiority of our model, e.g., the mIoU score under 5-shot setting on COCO-20^i is 5.6%+ better than previous state-of-the-arts. The code is available at https://github.com/Sam1224/SCCAN.
Video depth estimation aims to infer temporally consistent depth. Some methods achieve temporal consistency by finetuning a single-image depth model during test time using geometry and re-projection constraints, which is inefficient and not robust. An alternative approach is to learn how to enforce temporal consistency from data, but this requires well-designed models and sufficient video depth data. To address these challenges, we propose a plug-and-play framework called Neural Video Depth Stabilizer (NVDS) that stabilizes inconsistent depth estimations and can be applied to different single-image depth models without extra effort. We also introduce a large-scale dataset, Video Depth in the Wild (VDW), which consists of 14,203 videos with over two million frames, making it the largest natural-scene video depth dataset to our knowledge. We evaluate our method on the VDW dataset as well as two public benchmarks and demonstrate significant improvements in consistency, accuracy, and efficiency compared to previous approaches. Our work serves as a solid baseline and provides a data foundation for learning-based video depth models. We will release our dataset and code for future research.
3D semantic scene understanding tasks have achieved great success with the emergence of deep learning, but often require a huge amount of manually annotated training data. To alleviate the annotation cost, we propose the first weakly-supervised 3D instance segmentation method that only requires categorical semantic labels as supervision, and we do not need instance-level labels. The required semantic annotations can be either dense or extreme sparse (e.g. 0.02% of total points). Even without having any instance-related ground-truth, we design an approach to break point clouds into raw fragments and find the most confident samples for learning instance centroids. Furthermore, we construct a recomposed dataset using pseudo instances, which is used to learn our defined multilevel shape-aware objectness signal. An asymmetrical object inference algorithm is followed to process core points and boundary points with different strategies, and generate high-quality pseudo instance labels to guide iterative training. Experiments demonstrate that our method can achieve comparable results with recent fully supervised methods. By generating pseudo instance labels from categorical semantic labels, our designed approach can also assist existing methods for learning 3D instance segmentation at reduced annotation cost.