A surge of interest has emerged in utilizing Transformers in diverse vision tasks owing to its formidable performance. However, existing approaches primarily focus on optimizing internal model architecture designs that often entail significant trial and error with high burdens. In this work, we propose a new paradigm dubbed Decision Stream Calibration that boosts the performance of general Vision Transformers. To achieve this, we shed light on the information propagation mechanism in the learning procedure by exploring the correlation between different tokens and the relevance coefficient of multiple dimensions. Upon further analysis, it was discovered that 1) the final decision is associated with tokens of foreground targets, while token features of foreground target will be transmitted into the next layer as much as possible, and the useless token features of background area will be eliminated gradually in the forward propagation. 2) Each category is solely associated with specific sparse dimensions in the tokens. Based on the discoveries mentioned above, we designed a two-stage calibration scheme, namely ViT-Calibrator, including token propagation calibration stage and dimension propagation calibration stage. Extensive experiments on commonly used datasets show that the proposed approach can achieve promising results. The source codes are given in the supplements.
The problem of deep long-tailed learning, a prevalent challenge in the realm of generic visual recognition, persists in a multitude of real-world applications. To tackle the heavily-skewed dataset issue in long-tailed classification, prior efforts have sought to augment existing deep models with the elaborate class-balancing strategies, such as class rebalancing, data augmentation, and module improvement. Despite the encouraging performance, the limited class knowledge of the tailed classes in the training dataset still bottlenecks the performance of the existing deep models. In this paper, we propose an innovative long-tailed learning paradigm that breaks the bottleneck by guiding the learning of deep networks with external prior knowledge. This is specifically achieved by devising an elaborated ``prophetic'' teacher, termed as ``Propheter'', that aims to learn the potential class distributions. The target long-tailed prediction model is then optimized under the instruction of the well-trained ``Propheter'', such that the distributions of different classes are as distinguishable as possible from each other. Experiments on eight long-tailed benchmarks across three architectures demonstrate that the proposed prophetic paradigm acts as a promising solution to the challenge of limited class knowledge in long-tailed datasets. Our code and model can be found in the supplementary material.
The remarkable positive impact of Deep Neural Networks on many Artificial Intelligence (AI) tasks has led to the development of various high performance algorithms as well as specialized processors and accelerators. In this paper we address this scenario by demonstrating that the principles underlying the modern realization of the general matrix multiplication (GEMM) in conventional processor architectures, are also valid to achieve high performance for the type of operations that arise in deep learning (DL) on an exotic accelerator such as the AI Engine (AIE) tile embedded in Xilinx Versal platforms. In particular, our experimental results with a prototype implementation of the GEMM kernel, on a Xilinx Versal VCK190, delivers performance close to 86.7% of the theoretical peak that can be expected on an AIE tile, for 16-bit integer operands.
Considering the computation complexity, we propose a Guided Hybrid Quantization with One-to-one Self-Teaching (GHOST}) framework. More concretely, we first design a structure called guided quantization self-distillation (GQSD), which is an innovative idea for realizing lightweight through the synergy of quantization and distillation. The training process of the quantization model is guided by its full-precision model, which is time-saving and cost-saving without preparing a huge pre-trained model in advance. Second, we put forward a hybrid quantization (HQ) module to obtain the optimal bit width automatically under a constrained condition where a threshold for distribution distance between the center and samples is applied in the weight value search space. Third, in order to improve information transformation, we propose a one-to-one self-teaching (OST) module to give the student network a ability of self-judgment. A switch control machine (SCM) builds a bridge between the student network and teacher network in the same location to help the teacher to reduce wrong guidance and impart vital knowledge to the student. This distillation method allows a model to learn from itself and gain substantial improvement without any additional supervision. Extensive experiments on a multimodal dataset (VEDAI) and single-modality datasets (DOTA, NWPU, and DIOR) show that object detection based on GHOST outperforms the existing detectors. The tiny parameters (<9.7 MB) and Bit-Operations (BOPs) (<2158 G) compared with any remote sensing-based, lightweight or distillation-based algorithms demonstrate the superiority in the lightweight design domain. Our code and model will be released at https://github.com/icey-zhang/GHOST.
Vision transformers (ViTs) have achieved impressive results on various computer vision tasks in the last several years. In this work, we study the capability of frozen ViTs, pretrained only on visual data, to generalize to audio-visual data without finetuning any of its original parameters. To do so, we propose a latent audio-visual hybrid (LAVISH) adapter that adapts pretrained ViTs to audio-visual tasks by injecting a small number of trainable parameters into every layer of a frozen ViT. To efficiently fuse visual and audio cues, our LAVISH adapter uses a small set of latent tokens, which form an attention bottleneck, thus, eliminating the quadratic cost of standard cross-attention. Compared to the existing modality-specific audio-visual methods, our approach achieves competitive or even better performance on various audio-visual tasks while using fewer tunable parameters and without relying on costly audio pretraining or external audio encoders. Our code is available at https://genjib.github.io/project_page/LAVISH/
The last several years have witnessed remarkable progress in video-and-language (VidL) understanding. However, most modern VidL approaches use complex and specialized model architectures and sophisticated pretraining protocols, making the reproducibility, analysis and comparisons of these frameworks difficult. Hence, instead of proposing yet another new VidL model, this paper conducts a thorough empirical study demystifying the most important factors in the VidL model design. Among the factors that we investigate are (i) the spatiotemporal architecture design, (ii) the multimodal fusion schemes, (iii) the pretraining objectives, (iv) the choice of pretraining data, (v) pretraining and finetuning protocols, and (vi) dataset and model scaling. Our empirical study reveals that the most important design factors include: temporal modeling, video-to-text multimodal fusion, masked modeling objectives, and joint training on images and videos. Using these empirical insights, we then develop a step-by-step recipe, dubbed VindLU, for effective VidL pretraining. Our final model trained using our recipe achieves comparable or better than state-of-the-art results on several VidL tasks without relying on external CLIP pretraining. In particular, on the text-to-video retrieval task, our approach obtains 61.2% on DiDeMo, and 55.0% on ActivityNet, outperforming current SOTA by 7.8% and 6.1% respectively. Furthermore, our model also obtains state-of-the-art video question-answering results on ActivityNet-QA, MSRVTT-QA, MSRVTT-MC and TVQA. Our code and pretrained models are publicly available at: https://github.com/klauscc/VindLU.
We present Perceiver-VL, a vision-and-language framework that efficiently handles high-dimensional multimodal inputs such as long videos and text. Powered by the iterative latent cross-attention of Perceiver, our framework scales with linear complexity, in contrast to the quadratic complexity of self-attention used in many state-of-the-art transformer-based models. To further improve the efficiency of our framework, we also study applying LayerDrop on cross-attention layers and introduce a mixed-stream architecture for cross-modal retrieval. We evaluate Perceiver-VL on diverse video-text and image-text benchmarks, where Perceiver-VL achieves the lowest GFLOPs and latency while maintaining competitive performance. In addition, we also provide comprehensive analyses of various aspects of our framework, including pretraining data, scalability of latent size and input size, dropping cross-attention layers at inference to reduce latency, modality aggregation strategy, positional encoding, and weight initialization strategy. Our code and checkpoints are available at: https://github.com/zinengtang/Perceiver_VL
In this paper, we propose an accurate yet fast small object detection method for RSI, named SuperYOLO, which fuses multimodal data and performs high resolution (HR) object detection on multiscale objects by utilizing the assisted super resolution (SR) learning and considering both the detection accuracy and computation cost. First, we construct a compact baseline by removing the Focus module to keep the HR features and significantly overcomes the missing error of small objects. Second, we utilize pixel-level multimodal fusion (MF) to extract information from various data to facilitate more suitable and effective features for small objects in RSI. Furthermore, we design a simple and flexible SR branch to learn HR feature representations that can discriminate small objects from vast backgrounds with low-resolution (LR) input, thus further improving the detection accuracy. Moreover, to avoid introducing additional computation, the SR branch is discarded in the inference stage and the computation of the network model is reduced due to the LR input. Experimental results show that, on the widely used VEDAI RS dataset, SuperYOLO achieves an accuracy of 73.61% (in terms of mAP50), which is more than 10% higher than the SOTA large models such as YOLOv5l, YOLOv5x and RS designed YOLOrs. Meanwhile, the GFOLPs and parameter size of SuperYOLO are about 18.1x and 4.2x less than YOLOv5x. Our proposed model shows a favorable accuracy-speed trade-off compared to the state-of-art models. The code will be open sourced at https://github.com/icey-zhang/SuperYOLO.
Facial expression is an essential factor in conveying human emotional states and intentions. Although remarkable advancement has been made in facial expression recognition (FER) task, challenges due to large variations of expression patterns and unavoidable data uncertainties still remain. In this paper, we propose mid-level representation enhancement (MRE) and graph embedded uncertainty suppressing (GUS) addressing these issues. On one hand, MRE is introduced to avoid expression representation learning being dominated by a limited number of highly discriminative patterns. On the other hand, GUS is introduced to suppress the feature ambiguity in the representation space. The proposed method not only has stronger generalization capability to handle different variations of expression patterns but also more robustness to capture expression representations. Experimental evaluation on Aff-Wild2 have verified the effectiveness of the proposed method.
Training an effective video-and-language model intuitively requires multiple frames as model inputs. However, it is unclear whether using multiple frames is beneficial to downstream tasks, and if yes, whether the performance gain is worth the drastically-increased computation and memory costs resulting from using more frames. In this work, we explore single-frame models for video-and-language learning. On a diverse set of video-and-language tasks (including text-to-video retrieval and video question answering), we show the surprising result that, with large-scale pre-training and a proper frame ensemble strategy at inference time, a single-frame trained model that does not consider temporal information can achieve better performance than existing methods that use multiple frames for training. This result reveals the existence of a strong "static appearance bias" in popular video-and-language datasets. Therefore, to allow for a more comprehensive evaluation of video-and-language models, we propose two new retrieval tasks based on existing fine-grained action recognition datasets that encourage temporal modeling. Our code is available at https://github.com/jayleicn/singularity