In this paper, we for the first time explore helpful multi-modal contextual knowledge to understand novel categories for open-vocabulary object detection (OVD). The multi-modal contextual knowledge stands for the joint relationship across regions and words. However, it is challenging to incorporate such multi-modal contextual knowledge into OVD. The reason is that previous detection frameworks fail to jointly model multi-modal contextual knowledge, as object detectors only support vision inputs and no caption description is provided at test time. To this end, we propose a multi-modal contextual knowledge distillation framework, MMC-Det, to transfer the learned contextual knowledge from a teacher fusion transformer with diverse multi-modal masked language modeling (D-MLM) to a student detector. The diverse multi-modal masked language modeling is realized by an object divergence constraint upon traditional multi-modal masked language modeling (MLM), in order to extract fine-grained region-level visual contexts, which are vital to object detection. Extensive experiments performed upon various detection datasets show the effectiveness of our multi-modal context learning strategy, where our approach well outperforms the recent state-of-the-art methods.
Camouflaged object detection (COD) and camouflaged instance segmentation (CIS) aim to recognize and segment objects that are blended into their surroundings, respectively. While several deep neural network models have been proposed to tackle those tasks, augmentation methods for COD and CIS have not been thoroughly explored. Augmentation strategies can help improve models' performance by increasing the size and diversity of the training data and exposing the model to a wider range of variations in the data. Besides, we aim to automatically learn transformations that help to reveal the underlying structure of camouflaged objects and allow the model to learn to better identify and segment camouflaged objects. To achieve this, we propose a learnable augmentation method in the frequency domain for COD and CIS via the Fourier transform approach, dubbed CamoFA. Our method leverages a conditional generative adversarial network and cross-attention mechanism to generate a reference image and an adaptive hybrid swapping with parameters to mix the low-frequency component of the reference image and the high-frequency component of the input image. This approach aims to make camouflaged objects more visible for detection and segmentation models. Without bells and whistles, our proposed augmentation method boosts the performance of camouflaged object detectors and instance segmenters by large margins.
Mixture of experts (MoE) is a popular technique to improve capacity of Large Language Models (LLMs) with conditionally-activated parallel experts. However, serving MoE models on memory-constrained devices is challenging due to the large parameter size. Typical solutions such as memory swapping or expert pruning may lead to significantly higher latency or severe accuracy loss. In this paper, we introduce SwapMoE, a framework for efficient serving of MoE-based large language models with tunable memory budgets. The main idea of SwapMoE is to keep a small dynamic set of important experts, namely Virtual Experts, in the main memory for inference, while seamlessly maintaining how the Virtual Experts map to the actual experts. Experiments have shown that SwapMoE can reduce the memory footprint while maintaining reasonable accuracy. For example, on text summarization tasks with Switch Transformer, SwapMoE can reduce the memory consumption from 14.2 GiB to 4.7 GiB, together with 50\% latency reduction and a slight Rouge-2 score drop of 0.041.
Unlike cloud-based deep learning models that are often large and uniform, edge-deployed models usually demand customization for domain-specific tasks and resource-limited environments. Such customization processes can be costly and time-consuming due to the diversity of edge scenarios and the training load for each scenario. Although various approaches have been proposed for rapid resource-oriented customization and task-oriented customization respectively, achieving both of them at the same time is challenging. Drawing inspiration from the generative AI and the modular composability of neural networks, we introduce NN-Factory, an one-for-all framework to generate customized lightweight models for diverse edge scenarios. The key idea is to use a generative model to directly produce the customized models, instead of training them. The main components of NN-Factory include a modular supernet with pretrained modules that can be conditionally activated to accomplish different tasks and a generative module assembler that manipulate the modules according to task and sparsity requirements. Given an edge scenario, NN-Factory can efficiently customize a compact model specialized in the edge task while satisfying the edge resource constraints by searching for the optimal strategy to assemble the modules. Based on experiments on image classification and object detection tasks with different edge devices, NN-Factory is able to generate high-quality task- and resource-specific models within few seconds, faster than conventional model customization approaches by orders of magnitude.
The robustness of object detection models is a major concern when applied to real-world scenarios. The performance of most models tends to degrade when confronted with images affected by corruptions, since they are usually trained and evaluated on clean datasets. While numerous studies have explored the robustness of object detection models on natural images, there is a paucity of research focused on models applied to aerial images, which feature complex backgrounds, substantial variations in scales, and orientations of objects. This paper addresses the challenge of assessing the robustness of object detection models on aerial images, with a specific emphasis on scenarios where images are affected by clouds. In this study, we introduce two novel benchmarks based on DOTA-v1.0. The first benchmark encompasses 19 prevalent corruptions, while the second focuses on the cloud-corrupted condition-a phenomenon uncommon in natural images yet frequent in aerial photography. We systematically evaluate the robustness of mainstream object detection models and perform necessary ablation experiments. Through our investigations, we find that rotation-invariant modeling and enhanced backbone architectures can improve the robustness of models. Furthermore, increasing the capacity of Transformer-based backbones can strengthen their robustness. The benchmarks we propose and our comprehensive experimental analyses can facilitate research on robust object detection on aerial images. The codes and datasets are available at: https://github.com/hehaodong530/DOTA-C.
New 3+1D high-resolution radar sensors are gaining importance for 3D object detection in the automotive domain due to their relative affordability and improved detection compared to classic low-resolution radar sensors. One limitation of high-resolution radar sensors, compared to lidar sensors, is the sparsity of the generated point cloud. This sparsity could be partially overcome by accumulating radar point clouds of subsequent time steps. This contribution analyzes limitations of accumulating radar point clouds on the View-of-Delft dataset. By employing different ego-motion estimation approaches, the dataset's inherent constraints, and possible solutions are analyzed. Additionally, a learning-based instance motion estimation approach is deployed to investigate the influence of dynamic motion on the accumulated point cloud for object detection. Experiments document an improved object detection performance by applying an ego-motion estimation and dynamic motion correction approach.
Unsupervised domain adaptation (UDA) plays a crucial role in object detection when adapting a source-trained detector to a target domain without annotated data. In this paper, we propose a novel and effective four-step UDA approach that leverages self-supervision and trains source and target data concurrently. We harness self-supervised learning to mitigate the lack of ground truth in the target domain. Our method consists of the following steps: (1) identify the region with the highest-confidence set of detections in each target image, which serve as our pseudo-labels; (2) crop the identified region and generate a collection of its augmented versions; (3) combine these latter into a composite image; (4) adapt the network to the target domain using the composed image. Through extensive experiments under cross-camera, cross-weather, and synthetic-to-real scenarios, our approach achieves state-of-the-art performance, improving upon the nearest competitor by more than 2% in terms of mean Average Precision (mAP). The code is available at https://github.com/MohamedTEV/DACA.
Robust and efficient learning remains a challenging problem in robotics, in particular with complex visual inputs. Inspired by human attention mechanism, with which we quickly process complex visual scenes and react to changes in the environment, we think that embedding auxiliary information about focus point into robot learning would enhance efficiency and robustness of the learning process. In this paper, we propose a novel approach to model and emulate the human attention with an approximate prediction model. We then leverage this output and feed it as a structured auxiliary feature map into downstream learning tasks. We validate this idea by learning a prediction model from human-gaze recordings of manual driving in the real world. We test our approach on two learning tasks - object detection and imitation learning. Our experiments demonstrate that the inclusion of predicted human attention leads to improved robustness of the trained models to out-of-distribution samples and faster learning in low-data regime settings. Our work highlights the potential of incorporating structured auxiliary information in representation learning for robotics and opens up new avenues for research in this direction. All code and data are available online.
Video moment retrieval and highlight detection have received attention in the current era of video content proliferation, aiming to localize moments and estimate clip relevances based on user-specific queries. Given that the video content is continuous in time, there is often a lack of clear boundaries between temporal events in a video. This boundary ambiguity makes it challenging for the model to learn text-video clip correspondences, resulting in the subpar performance of existing methods in predicting target segments. To alleviate this problem, we propose to solve the two tasks jointly from the perspective of denoising generation. Moreover, the target boundary can be localized clearly by iterative refinement from coarse to fine. Specifically, a novel framework, DiffusionVMR, is proposed to redefine the two tasks as a unified conditional denoising generation process by combining the diffusion model. During training, Gaussian noise is added to corrupt the ground truth, with noisy candidates produced as input. The model is trained to reverse this noise addition process. In the inference phase, DiffusionVMR initiates directly from Gaussian noise and progressively refines the proposals from the noise to the meaningful output. Notably, the proposed DiffusionVMR inherits the advantages of diffusion models that allow for iteratively refined results during inference, enhancing the boundary transition from coarse to fine. Furthermore, the training and inference of DiffusionVMR are decoupled. An arbitrary setting can be used in DiffusionVMR during inference without consistency with the training phase. Extensive experiments conducted on five widely-used benchmarks (i.e., QVHighlight, Charades-STA, TACoS, YouTubeHighlights and TVSum) across two tasks (moment retrieval and/or highlight detection) demonstrate the effectiveness and flexibility of the proposed DiffusionVMR.
We present DySample, an ultra-lightweight and effective dynamic upsampler. While impressive performance gains have been witnessed from recent kernel-based dynamic upsamplers such as CARAFE, FADE, and SAPA, they introduce much workload, mostly due to the time-consuming dynamic convolution and the additional sub-network used to generate dynamic kernels. Further, the need for high-res feature guidance of FADE and SAPA somehow limits their application scenarios. To address these concerns, we bypass dynamic convolution and formulate upsampling from the perspective of point sampling, which is more resource-efficient and can be easily implemented with the standard built-in function in PyTorch. We first showcase a naive design, and then demonstrate how to strengthen its upsampling behavior step by step towards our new upsampler, DySample. Compared with former kernel-based dynamic upsamplers, DySample requires no customized CUDA package and has much fewer parameters, FLOPs, GPU memory, and latency. Besides the light-weight characteristics, DySample outperforms other upsamplers across five dense prediction tasks, including semantic segmentation, object detection, instance segmentation, panoptic segmentation, and monocular depth estimation. Code is available at https://github.com/tiny-smart/dysample.
Few-shot object detection aims to simultaneously localize and classify the objects in an image with limited training samples. However, most existing few-shot object detection methods focus on extracting the features of a few samples of novel classes that lack diversity. Hence, they may not be sufficient to capture the data distribution. To address that limitation, in this paper, we propose a novel approach in which we train a generator to generate synthetic data for novel classes. Still, directly training a generator on the novel class is not effective due to the lack of novel data. To overcome that issue, we leverage the large-scale dataset of base classes. Our overarching goal is to train a generator that captures the data variations of the base dataset. We then transform the captured variations into novel classes by generating synthetic data with the trained generator. To encourage the generator to capture data variations on base classes, we propose to train the generator with an optimal transport loss that minimizes the optimal transport distance between the distributions of real and synthetic data. Extensive experiments on two benchmark datasets demonstrate that the proposed method outperforms the state of the art. Source code will be available.
Since real-world training datasets cannot properly sample the long tail of the underlying data distribution, corner cases and rare out-of-domain samples can severely hinder the performance of state-of-the-art models. This problem becomes even more severe for dense tasks, such as 3D semantic segmentation, where points of non-standard objects can be confidently associated to the wrong class. In this work, we focus on improving the generalization to out-of-domain data. We achieve this by augmenting the training set with adversarial examples. First, we learn a set of vectors that deform the objects in an adversarial fashion. To prevent the adversarial examples from being too far from the existing data distribution, we preserve their plausibility through a series of constraints, ensuring sensor-awareness and shapes smoothness. Then, we perform adversarial augmentation by applying the learned sample-independent vectors to the available objects when training a model. We conduct extensive experiments across a variety of scenarios on data from KITTI, Waymo, and CrashD for 3D object detection, and on data from SemanticKITTI, Waymo, and nuScenes for 3D semantic segmentation. Despite training on a standard single dataset, our approach substantially improves the robustness and generalization of both 3D object detection and 3D semantic segmentation methods to out-of-domain data.
This paper proposes an efficient and effective method for joint gaze location detection (GL-D) and gaze object detection (GO-D), \emph{i.e.}, gaze following detection. Current approaches frame GL-D and GO-D as two separate tasks, employing a multi-stage framework where human head crops must first be detected and then be fed into a subsequent GL-D sub-network, which is further followed by an additional object detector for GO-D. In contrast, we reframe the gaze following detection task as detecting human head locations and their gaze followings simultaneously, aiming at jointly detect human gaze location and gaze object in a unified and single-stage pipeline. To this end, we propose GTR, short for \underline{G}aze following detection \underline{TR}ansformer, streamlining the gaze following detection pipeline by eliminating all additional components, leading to the first unified paradigm that unites GL-D and GO-D in a fully end-to-end manner. GTR enables an iterative interaction between holistic semantics and human head features through a hierarchical structure, inferring the relations of salient objects and human gaze from the global image context and resulting in an impressive accuracy. Concretely, GTR achieves a 12.1 mAP gain ($\mathbf{25.1}\%$) on GazeFollowing and a 18.2 mAP gain ($\mathbf{43.3\%}$) on VideoAttentionTarget for GL-D, as well as a 19 mAP improvement ($\mathbf{45.2\%}$) on GOO-Real for GO-D. Meanwhile, unlike existing systems detecting gaze following sequentially due to the need for a human head as input, GTR has the flexibility to comprehend any number of people's gaze followings simultaneously, resulting in high efficiency. Specifically, GTR introduces over a $\times 9$ improvement in FPS and the relative gap becomes more pronounced as the human number grows.
In the field of autonomous driving, accurate and comprehensive perception of the 3D environment is crucial. Bird's Eye View (BEV) based methods have emerged as a promising solution for 3D object detection using multi-view images as input. However, existing 3D object detection methods often ignore the physical context in the environment, such as sidewalk and vegetation, resulting in sub-optimal performance. In this paper, we propose a novel approach called SOGDet (Semantic-Occupancy Guided Multi-view 3D Object Detection), that leverages a 3D semantic-occupancy branch to improve the accuracy of 3D object detection. In particular, the physical context modeled by semantic occupancy helps the detector to perceive the scenes in a more holistic view. Our SOGDet is flexible to use and can be seamlessly integrated with most existing BEV-based methods. To evaluate its effectiveness, we apply this approach to several state-of-the-art baselines and conduct extensive experiments on the exclusive nuScenes dataset. Our results show that SOGDet consistently enhance the performance of three baseline methods in terms of nuScenes Detection Score (NDS) and mean Average Precision (mAP). This indicates that the combination of 3D object detection and 3D semantic occupancy leads to a more comprehensive perception of the 3D environment, thereby aiding build more robust autonomous driving systems. The codes are available at: https://github.com/zhouqiu/SOGDet.
Group regression is commonly used in 3D object detection to predict box parameters of similar classes in a joint head, aiming to benefit from similarities while separating highly dissimilar classes. For query-based perception methods, this has, so far, not been feasible. We close this gap and present a method to incorporate multi-class group regression, especially designed for the 3D domain in the context of autonomous driving, into existing attention and query-based perception approaches. We enhance a transformer based joint object detection and tracking model with this approach, and thoroughly evaluate its behavior and performance. For group regression, the classes of the nuScenes dataset are divided into six groups of similar shape and prevalence, each being regressed by a dedicated head. We show that the proposed method is applicable to many existing transformer based perception approaches and can bring potential benefits. The behavior of query group regression is thoroughly analyzed in comparison to a unified regression head, e.g. in terms of class-switching behavior and distribution of the output parameters. The proposed method offers many possibilities for further research, such as in the direction of deep multi-hypotheses tracking.
Knowledge distillation (KD) has shown potential for learning compact models in dense object detection. However, the commonly used softmax-based distillation ignores the absolute classification scores for individual categories. Thus, the optimum of the distillation loss does not necessarily lead to the optimal student classification scores for dense object detectors. This cross-task protocol inconsistency is critical, especially for dense object detectors, since the foreground categories are extremely imbalanced. To address the issue of protocol differences between distillation and classification, we propose a novel distillation method with cross-task consistent protocols, tailored for the dense object detection. For classification distillation, we address the cross-task protocol inconsistency problem by formulating the classification logit maps in both teacher and student models as multiple binary-classification maps and applying a binary-classification distillation loss to each map. For localization distillation, we design an IoU-based Localization Distillation Loss that is free from specific network structures and can be compared with existing localization distillation losses. Our proposed method is simple but effective, and experimental results demonstrate its superiority over existing methods. Code is available at https://github.com/TinyTigerPan/BCKD.
In panorama understanding, the widely used equirectangular projection (ERP) entails boundary discontinuity and spatial distortion. It severely deteriorates the conventional CNNs and vision Transformers on panoramas. In this paper, we propose a simple yet effective architecture named PanoSwin to learn panorama representations with ERP. To deal with the challenges brought by equirectangular projection, we explore a pano-style shift windowing scheme and novel pitch attention to address the boundary discontinuity and the spatial distortion, respectively. Besides, based on spherical distance and Cartesian coordinates, we adapt absolute positional embeddings and relative positional biases for panoramas to enhance panoramic geometry information. Realizing that planar image understanding might share some common knowledge with panorama understanding, we devise a novel two-stage learning framework to facilitate knowledge transfer from the planar images to panoramas. We conduct experiments against the state-of-the-art on various panoramic tasks, i.e., panoramic object detection, panoramic classification, and panoramic layout estimation. The experimental results demonstrate the effectiveness of PanoSwin in panorama understanding.
Object detection is a task that performs position identification and label classification of objects in images or videos. The information obtained through this process plays an essential role in various tasks in the field of computer vision. In object detection, the data utilized for training and validation typically originate from public datasets that are well-balanced in terms of the number of objects ascribed to each class in an image. However, in real-world scenarios, handling datasets with much greater class imbalance, i.e., very different numbers of objects for each class , is much more common, and this imbalance may reduce the performance of object detection when predicting unseen test images. In our study, thus, we propose a method that evenly distributes the classes in an image for training and validation, solving the class imbalance problem in object detection. Our proposed method aims to maintain a uniform class distribution through multi-label stratification. We tested our proposed method not only on public datasets that typically exhibit balanced class distribution but also on custom datasets that may have imbalanced class distribution. We found that our proposed method was more effective on datasets containing severe imbalance and less data. Our findings indicate that the proposed method can be effectively used on datasets with substantially imbalanced class distribution.
We examine how the choice of data-side attributes for two important visual tasks of image classification and object detection can aid in the choice or design of lightweight convolutional neural networks. We show by experimentation how four data attributes - number of classes, object color, image resolution, and object scale affect neural network model size and efficiency. Intra- and inter-class similarity metrics, based on metric learning, are defined to guide the evaluation of these attributes toward achieving lightweight models. Evaluations made using these metrics are shown to require 30x less computation than running full inference tests. We provide, as an example, applying the metrics and methods to choose a lightweight model for a robot path planning application and achieve computation reduction of 66% and accuracy gain of 3.5% over the pre-method model.
Vision Transformers achieve impressive accuracy across a range of visual recognition tasks. Unfortunately, their accuracy frequently comes with high computational costs. This is a particular issue in video recognition, where models are often applied repeatedly across frames or temporal chunks. In this work, we exploit temporal redundancy between subsequent inputs to reduce the cost of Transformers for video processing. We describe a method for identifying and re-processing only those tokens that have changed significantly over time. Our proposed family of models, Eventful Transformers, can be converted from existing Transformers (often without any re-training) and give adaptive control over the compute cost at runtime. We evaluate our method on large-scale datasets for video object detection (ImageNet VID) and action recognition (EPIC-Kitchens 100). Our approach leads to significant computational savings (on the order of 2-4x) with only minor reductions in accuracy.