LGViT: Dynamic Early Exiting for Accelerating Vision Transformer

Recently, the efficient deployment and acceleration of powerful vision transformers (ViTs) on resource-limited edge devices for providing multimedia services have become attractive tasks. Although early exiting is a feasible solution for accelerating inference, most works focus on convolutional neural networks (CNNs) and transformer models in natural language processing (NLP).Moreover, the direct application of early exiting methods to ViTs may result in substantial performance degradation. To tackle this challenge, we systematically investigate the efficacy of early exiting in ViTs and point out that the insufficient feature representations in shallow internal classifiers and the limited ability to capture target semantic information in deep internal classifiers restrict the performance of these methods. We then propose an early exiting framework for general ViTs termed LGViT, which incorporates heterogeneous exiting heads, namely, local perception head and global aggregation head, to achieve an efficiency-accuracy trade-off. In particular, we develop a novel two-stage training scheme, including end-to-end training and self-distillation with the backbone frozen to generate early exiting ViTs, which facilitates the fusion of global and local information extracted by the two types of heads. We conduct extensive experiments using three popular ViT backbones on three vision datasets. Results demonstrate that our LGViT can achieve competitive performance with approximately 1.8 $\times$ speed-up.

FHA-Kitchens: A Novel Dataset for Fine-Grained Hand Action Recognition in Kitchen Scenes

A typical task in the field of video understanding is hand action recognition, which has a wide range of applications. Existing works either mainly focus on full-body actions, or the defined action categories are relatively coarse-grained. In this paper, we propose FHA-Kitchens, a novel dataset of fine-grained hand actions in kitchen scenes. In particular, we focus on human hand interaction regions and perform deep excavation to further refine hand action information and interaction regions. Our FHA-Kitchens dataset consists of 2,377 video clips and 30,047 images collected from 8 different types of dishes, and all hand interaction regions in each image are labeled with high-quality fine-grained action classes and bounding boxes. We represent the action information in each hand interaction region as a triplet, resulting in a total of 878 action triplets. Based on the constructed dataset, we benchmark representative action recognition and detection models on the following three tracks: (1) supervised learning for hand interaction region and object detection, (2) supervised learning for fine-grained hand action recognition, and (3) intra- and inter-class domain generalization for hand interaction region detection. The experimental results offer compelling empirical evidence that highlights the challenges inherent in fine-grained hand action recognition, while also shedding light on potential avenues for future research, particularly in relation to pre-training strategy, model design, and domain generalization. The dataset will be released at https://github.com/tingZ123/FHA-Kitchens.

StructuredMesh: 3D Structured Optimization of Façade Components on Photogrammetric Mesh Models using Binary Integer Programming

The lack of fa\c{c}ade structures in photogrammetric mesh models renders them inadequate for meeting the demands of intricate applications. Moreover, these mesh models exhibit irregular surfaces with considerable geometric noise and texture quality imperfections, making the restoration of structures challenging. To address these shortcomings, we present StructuredMesh, a novel approach for reconstructing fa\c{c}ade structures conforming to the regularity of buildings within photogrammetric mesh models. Our method involves capturing multi-view color and depth images of the building model using a virtual camera and employing a deep learning object detection pipeline to semi-automatically extract the bounding boxes of fa\c{c}ade components such as windows, doors, and balconies from the color image. We then utilize the depth image to remap these boxes into 3D space, generating an initial fa\c{c}ade layout. Leveraging architectural knowledge, we apply binary integer programming (BIP) to optimize the 3D layout's structure, encompassing the positions, orientations, and sizes of all components. The refined layout subsequently informs fa\c{c}ade modeling through instance replacement. We conducted experiments utilizing building mesh models from three distinct datasets, demonstrating the adaptability, robustness, and noise resistance of our proposed methodology. Furthermore, our 3D layout evaluation metrics reveal that the optimized layout enhances precision, recall, and F-score by 6.5%, 4.5%, and 5.5%, respectively, in comparison to the initial layout.

GlyphControl: Glyph Conditional Control for Visual Text Generation

Recently, there has been a growing interest in developing diffusion-based text-to-image generative models capable of generating coherent and well-formed visual text. In this paper, we propose a novel and efficient approach called GlyphControl to address this task. Unlike existing methods that rely on character-aware text encoders like ByT5 and require retraining of text-to-image models, our approach leverages additional glyph conditional information to enhance the performance of the off-the-shelf Stable-Diffusion model in generating accurate visual text. By incorporating glyph instructions, users can customize the content, location, and size of the generated text according to their specific requirements. To facilitate further research in visual text generation, we construct a training benchmark dataset called LAION-Glyph. We evaluate the effectiveness of our approach by measuring OCR-based metrics and CLIP scores of the generated visual text. Our empirical evaluations demonstrate that GlyphControl outperforms the recent DeepFloyd IF approach in terms of OCR accuracy and CLIP scores, highlighting the efficacy of our method.

VanillaKD: Revisit the Power of Vanilla Knowledge Distillation from Small Scale to Large Scale

The tremendous success of large models trained on extensive datasets demonstrates that scale is a key ingredient in achieving superior results. Therefore, the reflection on the rationality of designing knowledge distillation (KD) approaches for limited-capacity architectures solely based on small-scale datasets is now deemed imperative. In this paper, we identify the \emph{small data pitfall} that presents in previous KD methods, which results in the underestimation of the power of vanilla KD framework on large-scale datasets such as ImageNet-1K. Specifically, we show that employing stronger data augmentation techniques and using larger datasets can directly decrease the gap between vanilla KD and other meticulously designed KD variants. This highlights the necessity of designing and evaluating KD approaches in the context of practical scenarios, casting off the limitations of small-scale datasets. Our investigation of the vanilla KD and its variants in more complex schemes, including stronger training strategies and different model capacities, demonstrates that vanilla KD is elegantly simple but astonishingly effective in large-scale scenarios. Without bells and whistles, we obtain state-of-the-art ResNet-50, ViT-S, and ConvNeXtV2-T models for ImageNet, which achieve 83.1\%, 84.3\%, and 85.0\% top-1 accuracy, respectively. PyTorch code and checkpoints can be found at https://github.com/Hao840/vanillaKD.

Improving Heterogeneous Model Reuse by Density Estimation

This paper studies multiparty learning, aiming to learn a model using the private data of different participants. Model reuse is a promising solution for multiparty learning, assuming that a local model has been trained for each party. Considering the potential sample selection bias among different parties, some heterogeneous model reuse approaches have been developed. However, although pre-trained local classifiers are utilized in these approaches, the characteristics of the local data are not well exploited. This motivates us to estimate the density of local data and design an auxiliary model together with the local classifiers for reuse. To address the scenarios where some local models are not well pre-trained, we further design a multiparty cross-entropy loss for calibration. Upon existing works, we address a challenging problem of heterogeneous model reuse from a decision theory perspective and take advantage of recent advances in density estimation. Experimental results on both synthetic and benchmark data demonstrate the superiority of the proposed method.

EfficientViT: Memory Efficient Vision Transformer with Cascaded Group Attention

Vision transformers have shown great success due to their high model capabilities. However, their remarkable performance is accompanied by heavy computation costs, which makes them unsuitable for real-time applications. In this paper, we propose a family of high-speed vision transformers named EfficientViT. We find that the speed of existing transformer models is commonly bounded by memory inefficient operations, especially the tensor reshaping and element-wise functions in MHSA. Therefore, we design a new building block with a sandwich layout, i.e., using a single memory-bound MHSA between efficient FFN layers, which improves memory efficiency while enhancing channel communication. Moreover, we discover that the attention maps share high similarities across heads, leading to computational redundancy. To address this, we present a cascaded group attention module feeding attention heads with different splits of the full feature, which not only saves computation cost but also improves attention diversity. Comprehensive experiments demonstrate EfficientViT outperforms existing efficient models, striking a good trade-off between speed and accuracy. For instance, our EfficientViT-M5 surpasses MobileNetV3-Large by 1.9% in accuracy, while getting 40.4% and 45.2% higher throughput on Nvidia V100 GPU and Intel Xeon CPU, respectively. Compared to the recent efficient model MobileViT-XXS, EfficientViT-M2 achieves 1.8% superior accuracy, while running 5.8x/3.7x faster on the GPU/CPU, and 7.4x faster when converted to ONNX format. Code and models are available at https://github.com/microsoft/Cream/tree/main/EfficientViT.

SeqTrack: Sequence to Sequence Learning for Visual Object Tracking

In this paper, we present a new sequence-to-sequence learning framework for visual tracking, dubbed SeqTrack. It casts visual tracking as a sequence generation problem, which predicts object bounding boxes in an autoregressive fashion. This is different from prior Siamese trackers and transformer trackers, which rely on designing complicated head networks, such as classification and regression heads. SeqTrack only adopts a simple encoder-decoder transformer architecture. The encoder extracts visual features with a bidirectional transformer, while the decoder generates a sequence of bounding box values autoregressively with a causal transformer. The loss function is a plain cross-entropy. Such a sequence learning paradigm not only simplifies tracking framework, but also achieves competitive performance on benchmarks. For instance, SeqTrack gets 72.5% AUC on LaSOT, establishing a new state-of-the-art performance. Code and models are available at here.

Implicit Temporal Modeling with Learnable Alignment for Video Recognition

Contrastive language-image pretraining (CLIP) has demonstrated remarkable success in various image tasks. However, how to extend CLIP with effective temporal modeling is still an open and crucial problem. Existing factorized or joint spatial-temporal modeling trades off between the efficiency and performance. While modeling temporal information within straight through tube is widely adopted in literature, we find that simple frame alignment already provides enough essence without temporal attention. To this end, in this paper, we proposed a novel Implicit Learnable Alignment (ILA) method, which minimizes the temporal modeling effort while achieving incredibly high performance. Specifically, for a frame pair, an interactive point is predicted in each frame, serving as a mutual information rich region. By enhancing the features around the interactive point, two frames are implicitly aligned. The aligned features are then pooled into a single token, which is leveraged in the subsequent spatial self-attention. Our method allows eliminating the costly or insufficient temporal self-attention in video. Extensive experiments on benchmarks demonstrate the superiority and generality of our module. Particularly, the proposed ILA achieves a top-1 accuracy of 88.7% on Kinetics-400 with much fewer FLOPs compared with Swin-L and ViViT-H. Code is released at https://github.com/Francis-Rings/ILA .

BCE-Net: Reliable Building Footprints Change Extraction based on Historical Map and Up-to-Date Images using Contrastive Learning

Automatic and periodic recompiling of building databases with up-to-date high-resolution images has become a critical requirement for rapidly developing urban environments. However, the architecture of most existing approaches for change extraction attempts to learn features related to changes but ignores objectives related to buildings. This inevitably leads to the generation of significant pseudo-changes, due to factors such as seasonal changes in images and the inclination of building fa\c{c}ades. To alleviate the above-mentioned problems, we developed a contrastive learning approach by validating historical building footprints against single up-to-date remotely sensed images. This contrastive learning strategy allowed us to inject the semantics of buildings into a pipeline for the detection of changes, which is achieved by increasing the distinguishability of features of buildings from those of non-buildings. In addition, to reduce the effects of inconsistencies between historical building polygons and buildings in up-to-date images, we employed a deformable convolutional neural network to learn offsets intuitively. In summary, we formulated a multi-branch building extraction method that identifies newly constructed and removed buildings, respectively. To validate our method, we conducted comparative experiments using the public Wuhan University building change detection dataset and a more practical dataset named SI-BU that we established. Our method achieved F1 scores of 93.99% and 70.74% on the above datasets, respectively. Moreover, when the data of the public dataset were divided in the same manner as in previous related studies, our method achieved an F1 score of 94.63%, which surpasses that of the state-of-the-art method.