Extracting robust feature representation is one of the key challenges in object re-identification (ReID). Although convolution neural network (CNN)-based methods have achieved great success, they only process local neighborhood at a time and suffer from information loss on details caused by downsampling operators (e.g. pooling strided convolution). To overcome these limitations, we propose pure transformer-based ReID framework named TransReID. Specifically, first encode an image as sequence...

Locating 3D objects from a single RGB image via Perspective-n-Points (PnP) is long-standing problem in computer vision. Driven by end-to-end deep learning, recent studies suggest interpreting PnP as differentiable layer, so that 2D-3D point correspondences can be partly learned backpropagating the gradient w.r.t. object pose. Yet, learning entire set of unrestricted points scratch fails to converge with existing approaches, since deterministic pose inherently non-differentiable. In this...

Existing semantic segmentation works have been mainly focused on designing effective decoders; however, the computational load introduced by overall structure has long ignored, which hinders their applications resource-constrained hardwares. In this paper, we propose a head-free lightweight architecture specifically for segmentation, named Adaptive Frequency Transformer (AFFormer). AFFormer adopts parallel to leverage prototype representations as specific learnable local descriptions...

To build a high-quality open-domain chatbot, we introduce the effective training process of PLATO-2 via curriculum learning.There are two stages involved in learning process.In first stage, coarse-grained generation model is trained to learn response under simplified framework oneto-one mapping.In second finegrained generative augmented with latent variables and an evaluation further generate diverse responses select best response, respectively.PLATO-2 was on both Chinese English data, whose...

Person Re-identification (ReID) plays a more and crucial role in recent years with wide range of applications. Existing ReID methods are suffering from the challenges misalignment occlusions, which degrade performance dramatically. Most tackle such by utilizing external tools to locate body parts or exploiting matching strategies. Nevertheless, inevitable domain gap between datasets utilized for complicated process make these unreliable sensitive noises. In this paper, we propose Region...

Locating 3D objects from a single RGB image via Perspective-n-Point (PnP) is long-standing problem in computer vision. Driven by end-to-end deep learning, recent studies suggest interpreting PnP as differentiable layer, allowing for partial learning of 2D-3D point correspondences backpropagating the gradients pose loss. Yet, entire scratch highly challenging, particularly ambiguous solutions, where globally optimal theoretically non-differentiable w.r.t. points. In this paper, we propose...

We developed a large multiplexing capacity dense ultra-short (DUS)-FBG array for high spatial resolution distributed sensing applications. With identical central wavelength and low peak reflectivity (−40 dB), all the FBGs share short length (1 mm) extremely small spacing (500 μm). The lower crosstalk of DUS-FBG is investigated through both simulation experiment. Use interrogated by optical frequency domain reflectometry (OFDR) temperature non-uniform strain was conducted. demonstrated over...

Over the past two decades, traditional block-based video coding has made remarkable progress and spawned a series of well-known standards such as MPEG-4, H.264/AVC H.265/HEVC. On other hand, deep neural networks (DNNs) have shown their powerful capacity for visual content understanding, feature extraction compact representation. Some previous works explored learnt algorithms in an end-to-end manner, which show great potential compared with methods. In this paper, we propose framework (NVC),...

Extracting robust feature representation is one of the key challenges in object re-identification (ReID). Although convolution neural network (CNN)-based methods have achieved great success, they only process local neighborhood at a time and suffer from information loss on details caused by downsampling operators (e.g. pooling strided convolution). To overcome these limitations, we propose pure transformer-based ReID framework named TransReID. Specifically, first encode an image as sequence...

Vision transformers (ViTs) have been an alternative design paradigm to convolutional neural networks (CNNs). However, the training of ViTs is much harder than CNNs, as it sensitive parameters, such learning rate, optimizer and warmup epoch. The reasons for difficulty are empirically analysed in paper Early Convolutions Help Transformers See Better, authors conjecture that issue lies with patchify-stem ViT models. In this paper, we further investigate problem extend above conclusion: only...

Text-to-3D generation has recently garnered significant attention, fueled by 2D diffusion models trained on billions of image-text pairs. Existing methods primarily rely score distillation to leverage the priors supervise 3D models, e.g., NeRF. However, is prone suffer view inconsistency problem, and implicit NeRF modeling can also lead an arbitrary shape, thus leading less realistic uncontrollable generation. In this work, we propose a flexible framework Points-to-3D bridge gap between...

The quadratic computational complexity to the number of tokens limits practical applications Vision Transformers (ViTs). Several works propose prune redundant achieve efficient ViTs. However, these methods generally suffer from (i) dramatic accuracy drops, (ii) application difficulty in local vision transformer, and (iii) non-general-purpose networks for downstream tasks. In this work, we a novel Semantic Token ViT (STViT), global transformers, which can also be revised serve as backbone...

Image-text retrieval is a central problem for understanding the semantic relationship between vision and language, serves as basis various visual language tasks. Most previous works either simply learn coarse-grained representations of overall image text, or elaborately establish correspondence regions pixels text words. However, close relations coarse- fine-grained each modality are important image-text but almost neglected. As result, such inevitably suffer from low accuracy heavy...

This paper explores the application of Orff music concept in mass entertainment. is an artistic form that holistic, natural and closely connected to life. Its elemental nature encompasses multiple elements, breaking excessive emphasis on traditional musical elements returning expression perception. When applied entertainment experience, it has advantages being simple operate highly improvisational, enabling people different age groups participate easily enhancing interactive cooperation....

With the advancement of geo-systems and increased availability satellite data, a plethora Land-Use Land-Cover (LULC) products have been developed. The existing LULC primarily relied on time-series imagery to classify land by pixel-based classifiers, allowing for local analysis accurate boundary detection. However, advent deep learning has shifted towards use patch-based CNN models generating cover maps. In this paper, (1) we create training dataset China using voting strategy based three...

Current research on cross-modal retrieval is mostly English-oriented, as the availability of a large number English-oriented human-labeled vision-language corpora. In order to break limit non-English labeled data, cross-lingual (CCR) has attracted increasing attention. Most CCR methods construct pseudo-parallel corpora via Machine Translation (MT) achieve transfer. However, translated sentences from MT are generally imperfect in describing corresponding visual contents. Improperly assuming...

Video object detection (VOD) has been a rising topic in recent years due to the challenges such as occlusion, motion blur, etc. To deal with these challenges, feature aggregation from local or global support frames is verified effective. exploit better aggregation, this paper, we propose two improvements over previous works: class-constrained spatial-temporal relation network and correlation-based alignment module. For class constrained network, it operates on region proposals, learns kinds...

Instance segmentation is an accurate and reliable method to segment adhesive pigs’ images, critical for providing health welfare information on individual pigs, such as body condition score, live weight, activity behaviors in group-housed pig environments. In this paper, a PigMS R-CNN framework based mask scoring (MS R-CNN) explored areas group-pig separate the identification location of pigs. The consists three processes. First, residual network 101-layers, combined with feature pyramid...

Motion recognition is a promising direction in computer vision, but the training of video classification models much harder than images due to insufficient data and considerable parameters. To get around this, some works strive explore multimodal cues from RGB-D data. Although improving motion extent, these methods still face sub-optimal situations following aspects: (i) Data augmentation, i.e., scale datasets limited, few efforts have been made novel augmentation strategies for videos; (ii)...

This paper introduces Amazon Robotic Manipulation Benchmark (ARMBench), a large-scale, object-centric benchmark dataset for robotic manipulation in the context of warehouse. Automation operations modern warehouses requires manipulator to deal with wide variety objects, unstructured storage, and dynamically changing inventory. Such settings pose challenges perceiving identity, physical characteristics, state objects during manipulation. Existing datasets consider limited set or utilize 3D...