Single image depth estimation is a foundational task in computer vision and generative modeling. However, prevailing depth estimation models grapple with accommodating the increasing resolutions commonplace in today's consumer cameras and devices. Existing high-resolution strategies show promise, but they often face limitations, ranging from error propagation to the loss of high-frequency details. We present PatchFusion, a novel tile-based framework with three key components to improve the current state of the art: (1) A patch-wise fusion network that fuses a globally-consistent coarse prediction with finer, inconsistent tiled predictions via high-level feature guidance, (2) A Global-to-Local (G2L) module that adds vital context to the fusion network, discarding the need for patch selection heuristics, and (3) A Consistency-Aware Training (CAT) and Inference (CAI) approach, emphasizing patch overlap consistency and thereby eradicating the necessity for post-processing. Experiments on UnrealStereo4K, MVS-Synth, and Middleburry 2014 demonstrate that our framework can generate high-resolution depth maps with intricate details. PatchFusion is independent of the base model for depth estimation. Notably, our framework built on top of SOTA ZoeDepth brings improvements for a total of 17.3% and 29.4% in terms of the root mean squared error (RMSE) on UnrealStereo4K and MVS-Synth, respectively.
Exploring open-vocabulary video action recognition is a promising venture, which aims to recognize previously unseen actions within any arbitrary set of categories. Existing methods typically adapt pretrained image-text models to the video domain, capitalizing on their inherent strengths in generalization. A common thread among such methods is the augmentation of visual embeddings with temporal information to improve the recognition of seen actions. Yet, they compromise with standard less-informative action descriptions, thus faltering when confronted with novel actions. Drawing inspiration from human cognitive processes, we argue that augmenting text embeddings with human prior knowledge is pivotal for open-vocabulary video action recognition. To realize this, we innovatively blend video models with Large Language Models (LLMs) to devise Action-conditioned Prompts. Specifically, we harness the knowledge in LLMs to produce a set of descriptive sentences that contain distinctive features for identifying given actions. Building upon this foundation, we further introduce a multi-modal action knowledge alignment mechanism to align concepts in video and textual knowledge encapsulated within the prompts. Extensive experiments on various video benchmarks, including zero-shot, few-shot, and base-to-novel generalization settings, demonstrate that our method not only sets new SOTA performance but also possesses excellent interpretability.
Predicting typhoon intensity accurately across space and time is crucial for issuing timely disaster warnings and facilitating emergency response. This has vast potential for minimizing life losses and property damages as well as reducing economic and environmental impacts. Leveraging satellite imagery for scenario analysis is effective but also introduces additional challenges due to the complex relations among clouds and the highly dynamic context. Existing deep learning methods in this domain rely on convolutional neural networks (CNNs), which suffer from limited per-layer receptive fields. This limitation hinders their ability to capture long-range dependencies and global contextual knowledge during inference. In response, we introduce a novel approach, namely "Typhoon Intensity Transformer" (Tint), which leverages self-attention mechanisms with global receptive fields per layer. Tint adopts a sequence-to-sequence feature representation learning perspective. It begins by cutting a given satellite image into a sequence of patches and recursively employs self-attention operations to extract both local and global contextual relations between all patch pairs simultaneously, thereby enhancing per-patch feature representation learning. Extensive experiments on a publicly available typhoon benchmark validate the efficacy of Tint in comparison with both state-of-the-art deep learning and conventional meteorological methods. Our code is available at https://github.com/chen-huanxin/Tint.
Editing visual content on videos remains a formidable challenge with two main issues: 1) direct and easy user control to produce 2) natural editing results without unsightly distortion and artifacts after changing shape, expression and layout. Inspired by DragGAN, a recent image-based drag-style editing technique, we address above issues by proposing DragVideo, where a similar drag-style user interaction is adopted to edit video content while maintaining temporal consistency. Empowered by recent diffusion models as in DragDiffusion, DragVideo contains the novel Drag-on-Video U-Net (DoVe) editing method, which optimizes diffused video latents generated by video U-Net to achieve the desired control. Specifically, we use Sample-specific LoRA fine-tuning and Mutual Self-Attention control to ensure faithful reconstruction of video from the DoVe method. We also present a series of testing examples for drag-style video editing and conduct extensive experiments across a wide array of challenging editing tasks, such as motion editing, skeleton editing, etc, underscoring DragVideo's versatility and generality. Our codes including the DragVideo web user interface will be released.
Recent studies have demonstrated the effectiveness of Gated Linear Units (GLU) in enhancing transformer models, particularly in Large Language Models (LLMs). Additionally, utilizing a parallel configuration within each Transformer block rather than the conventional serialized method has been revealed to accelerate the training of LLMs without significantly impacting performance. However, when the MLP and attention block were run in parallel for the image classification task, we observed a noticeable decline in performance. We propose a novel transformer variant that integrates non-linearity within the attention block to tackle this problem. We implemented the GLU-based activation function on the Value tensor, and this new technique surpasses the current state-of-the-art S/16 variant of Vision Transformers by 0.6% on the ImageNet-1K dataset while utilizing fewer parameters. It also supersedes the B/16 variant while using only half the parameters. Furthermore, we provide results with the GELU activation function variant to confirm our assertions. Lastly, we showcase that the MABViT variants exhibit greater potential when utilized in deep transformers compared to the standard architecture.
Unsupervised representation learning aims at finding methods that learn representations from data without annotation-based signals. Abstaining from annotations not only leads to economic benefits but may - and to some extent already does - result in advantages regarding the representation's structure, robustness, and generalizability to different tasks. In the long run, unsupervised methods are expected to surpass their supervised counterparts due to the reduction of human intervention and the inherently more general setup that does not bias the optimization towards an objective originating from specific annotation-based signals. While major advantages of unsupervised representation learning have been recently observed in natural language processing, supervised methods still dominate in vision domains for most tasks. In this dissertation, we contribute to the field of unsupervised (visual) representation learning from three perspectives: (i) Learning representations: We design unsupervised, backpropagation-free Convolutional Self-Organizing Neural Networks (CSNNs) that utilize self-organization- and Hebbian-based learning rules to learn convolutional kernels and masks to achieve deeper backpropagation-free models. (ii) Evaluating representations: We build upon the widely used (non-)linear evaluation protocol to define pretext- and target-objective-independent metrics for measuring and investigating the objective function mismatch between various unsupervised pretext tasks and target tasks. (iii) Transferring representations: We contribute CARLANE, the first 3-way sim-to-real domain adaptation benchmark for 2D lane detection, and a method based on prototypical self-supervised learning. Finally, we contribute a content-consistent unpaired image-to-image translation method that utilizes masks, global and local discriminators, and similarity sampling to mitigate content inconsistencies.
Parkinson's Disease (PD) is a neurodegenerative neurological disorder that impacts movement and afflicts over 10 million people worldwide. Previous researches have come up with deep learning models for predicting Parkinson's disease primarily using medical images and didn't leverage the manifold structure in the dataset. Our study introduces a multimodal approach with both image and non-image features with a contrastive cross-view graph fusion for Parkinson's disease classification. Specifically, we designed a multimodal co-attention module to integrate embeddings from two distinct graph views derived from low dimensional representation of images and clinical features, enabling the extraction of more stable and structured features from the multiview data. Additionally, we have devised a simplified fusion method utilizing a contrastive loss for positive and negative pairs, to enhance the model's overall cross-view fusion learning capabilities. In our experiments, the graph-view multimodal approach can achieve an accuracy rate of 91% and an AUC of 92.8% in five-fold cross-validation, and it also demonstrates superior predictive capabilities on non-image data as compared to methods that rely solely on machine learning methods.
There are not many large medical image datasets available. For these datasets, too small deep learning models can't learn useful features, so they don't work well due to underfitting, and too big models tend to overfit the limited data. As a result, there is a compromise between the two issues. This paper proposes a training strategy Medi-CAT to overcome the underfitting and overfitting phenomena in medical imaging datasets. Specifically, the proposed training methodology employs large pre-trained vision transformers to overcome underfitting and adversarial and contrastive learning techniques to prevent overfitting. The proposed method is trained and evaluated on four medical image classification datasets from the MedMNIST collection. Our experimental results indicate that the proposed approach improves the accuracy up to 2% on three benchmark datasets compared to well-known approaches, whereas it increases the performance up to 4.1% over the baseline methods.
Cross-domain few-shot hyperspectral image classification focuses on learning prior knowledge from a large number of labeled samples from source domain and then transferring the knowledge to the tasks which contain only few labeled samples in target domains. Following the metric-based manner, many current methods first extract the features of the query and support samples, and then directly predict the classes of query samples according to their distance to the support samples or prototypes. The relations between samples have not been fully explored and utilized. Different from current works, this paper proposes to learn sample relations from different views and take them into the model learning process, to improve the cross-domain few-shot hyperspectral image classification. Building on current DCFSL method which adopts a domain discriminator to deal with domain-level distribution difference, the proposed method applys contrastive learning to learn the class-level sample relations to obtain more discriminable sample features. In addition, it adopts a transformer based cross-attention learning module to learn the set-level sample relations and acquire the attentions from query samples to support samples. Our experimental results have demonstrated the contribution of the multi-view relation learning mechanism for few-shot hyperspectral image classification when compared with the state of the art methods.
Detecting machine malfunctions at an early stage is crucial for reducing interruptions in operational processes within industrial settings. Recently, the deep learning approach has started to be preferred for the detection of failures in machines. Deep learning provides an effective solution in fault detection processes thanks to automatic feature extraction. In this study, a deep learning-based system was designed to analyze the sound signals produced by industrial machines. Acoustic sound signals were converted into Mel spectrograms. For the purpose of classifying spectrogram images, the DenseNet-169 model, a deep learning architecture recognized for its effectiveness in image classification tasks, was used. The model was trained using the transfer learning method on the MIMII dataset including sounds from four types of industrial machines. The results showed that the proposed method reached an accuracy rate varying between 97.17% and 99.87% at different Sound Noise Rate levels.