Enhancing Visual Continual Learning with Language-Guided Supervision

Continual learning (CL) aims to empower models to learn new tasks without forgetting previously acquired knowledge. Most prior works concentrate on the techniques of architectures, replay data, regularization, \etc. However, the category name of each class is largely neglected. Existing methods commonly utilize the one-hot labels and randomly initialize the classifier head. We argue that the scarce semantic information conveyed by the one-hot labels hampers the effective knowledge transfer across tasks. In this paper, we revisit the role of the classifier head within the CL paradigm and replace the classifier with semantic knowledge from pretrained language models (PLMs). Specifically, we use PLMs to generate semantic targets for each class, which are frozen and serve as supervision signals during training. Such targets fully consider the semantic correlation between all classes across tasks. Empirical studies show that our approach mitigates forgetting by alleviating representation drifting and facilitating knowledge transfer across tasks. The proposed method is simple to implement and can seamlessly be plugged into existing methods with negligible adjustments. Extensive experiments based on eleven mainstream baselines demonstrate the effectiveness and generalizability of our approach to various protocols. For example, under the class-incremental learning setting on ImageNet-100, our method significantly improves the Top-1 accuracy by 3.2\% to 6.1\% while reducing the forgetting rate by 2.6\% to 13.1\%.

Continual Forgetting for Pre-trained Vision Models

For privacy and security concerns, the need to erase unwanted information from pre-trained vision models is becoming evident nowadays. In real-world scenarios, erasure requests originate at any time from both users and model owners. These requests usually form a sequence. Therefore, under such a setting, selective information is expected to be continuously removed from a pre-trained model while maintaining the rest. We define this problem as continual forgetting and identify two key challenges. (i) For unwanted knowledge, efficient and effective deleting is crucial. (ii) For remaining knowledge, the impact brought by the forgetting procedure should be minimal. To address them, we propose Group Sparse LoRA (GS-LoRA). Specifically, towards (i), we use LoRA modules to fine-tune the FFN layers in Transformer blocks for each forgetting task independently, and towards (ii), a simple group sparse regularization is adopted, enabling automatic selection of specific LoRA groups and zeroing out the others. GS-LoRA is effective, parameter-efficient, data-efficient, and easy to implement. We conduct extensive experiments on face recognition, object detection and image classification and demonstrate that GS-LoRA manages to forget specific classes with minimal impact on other classes. Codes will be released on \url{https://github.com/bjzhb666/GS-LoRA}.

MVPatch: More Vivid Patch for Adversarial Camouflaged Attacks on Object Detectors in the Physical World

Recent investigations demonstrate that adversarial patches can be utilized to manipulate the result of object detection models. However, the conspicuous patterns on these patches may draw more attention and raise suspicions among humans. Moreover, existing works have primarily focused on enhancing the efficacy of attacks in the physical domain, rather than seeking to optimize their stealth attributes and transferability potential. To address these issues, we introduce a dual-perception-based attack framework that generates an adversarial patch known as the More Vivid Patch (MVPatch). The framework consists of a model-perception degradation method and a human-perception improvement method. To derive the MVPatch, we formulate an iterative process that simultaneously constrains the efficacy of multiple object detectors and refines the visual correlation between the generated adversarial patch and a realistic image. Our method employs a model-perception-based approach that reduces the object confidence scores of several object detectors to boost the transferability of adversarial patches. Further, within the human-perception-based framework, we put forward a lightweight technique for visual similarity measurement that facilitates the development of inconspicuous and natural adversarial patches and eliminates the reliance on additional generative models. Additionally, we introduce the naturalness score and transferability score as metrics for an unbiased assessment of various adversarial patches' natural appearance and transferability capacity. Extensive experiments demonstrate that the proposed MVPatch algorithm achieves superior attack transferability compared to similar algorithms in both digital and physical domains while also exhibiting a more natural appearance. These findings emphasize the remarkable stealthiness and transferability of the proposed MVPatch attack algorithm.

Plane Constraints Aided Multi-Vehicle Cooperative Positioning Using Factor Graph Optimization

The development of vehicle-to-vehicle (V2V) communication facil-itates the study of cooperative positioning (CP) techniques for vehicular applications. The CP methods can improve the posi-tioning availability and accuracy by inter-vehicle ranging and data exchange between vehicles. However, the inter-vehicle rang-ing can be easily interrupted due to many factors such as obsta-cles in-between two cars. Without inter-vehicle ranging, the other cooperative data such as vehicle positions will be wasted, leading to performance degradation of range-based CP methods. To fully utilize the cooperative data and mitigate the impact of inter-vehicle ranging loss, a novel cooperative positioning method aided by plane constraints is proposed in this paper. The positioning results received from cooperative vehicles are used to construct the road plane for each vehicle. The plane parameters are then introduced into CP scheme to impose constraints on positioning solutions. The state-of-art factor graph optimization (FGO) algo-rithm is employed to integrate the plane constraints with raw data of Global Navigation Satellite Systems (GNSS) as well as inter-vehicle ranging measurements. The proposed CP method has the ability to resist the interruptions of inter-vehicle ranging since the plane constraints are computed by just using position-related data. A vehicle can still benefit from the position data of cooperative vehicles even if the inter-vehicle ranging is unavaila-ble. The experimental results indicate the superiority of the pro-posed CP method in positioning performance over the existing methods, especially when the inter-ranging interruptions occur.

Dive Deeper into Rectifying Homography for Stereo Camera Online Self-Calibration

Accurate estimation of stereo camera extrinsic parameters is the key to guarantee the performance of stereo matching algorithms. In prior arts, the online self-calibration of stereo cameras has commonly been formulated as a specialized visual odometry problem, without taking into account the principles of stereo rectification. In this paper, we first delve deeply into the concept of rectifying homography, which serves as the cornerstone for the development of our novel stereo camera online self-calibration algorithm, for cases where only a single pair of images is available. Furthermore, we introduce a simple yet effective solution for global optimum extrinsic parameter estimation in the presence of stereo video sequences. Additionally, we emphasize the impracticality of using three Euler angles and three components in the translation vectors for performance quantification. Instead, we introduce four new evaluation metrics to quantify the robustness and accuracy of extrinsic parameter estimation, applicable to both single-pair and multi-pair cases. Extensive experiments conducted across indoor and outdoor environments using various experimental setups validate the effectiveness of our proposed algorithm. The comprehensive evaluation results demonstrate its superior performance in comparison to the baseline algorithm. Our source code, demo video, and supplement are publicly available at mias.group/StereoCalibrator.

Self-Training Guided Disentangled Adaptation for Cross-Domain Remote Sensing Image Semantic Segmentation

Deep convolutional neural networks (DCNNs) based remote sensing (RS) image semantic segmentation technology has achieved great success used in many real-world applications such as geographic element analysis. However, strong dependency on annotated data of specific scene makes it hard for DCNNs to fit different RS scenes. To solve this problem, recent works gradually focus on cross-domain RS image semantic segmentation task. In this task, different ground sampling distance, remote sensing sensor variation and different geographical landscapes are three main factors causing dramatic domain shift between source and target images. To decrease the negative influence of domain shift, we propose a self-training guided disentangled adaptation network (ST-DASegNet). We first propose source student backbone and target student backbone to respectively extract the source-style and target-style feature for both source and target images. Towards the intermediate output feature maps of each backbone, we adopt adversarial learning for alignment. Then, we propose a domain disentangled module to extract the universal feature and purify the distinct feature of source-style and target-style features. Finally, these two features are fused and served as input of source student decoder and target student decoder to generate final predictions. Based on our proposed domain disentangled module, we further propose exponential moving average (EMA) based cross-domain separated self-training mechanism to ease the instability and disadvantageous effect during adversarial optimization. Extensive experiments and analysis on benchmark RS datasets show that ST-DASegNet outperforms previous methods on cross-domain RS image semantic segmentation task and achieves state-of-the-art (SOTA) results. Our code is available at https://github.com/cv516Buaa/ST-DASegNet.

PGGANet: Pose Guided Graph Attention Network for Person Re-identification

Person re-identification (ReID) aims at retrieving a person from images captured by different cameras. For deep-learning-based ReID methods, it has been proved that using local features together with global feature of person image could help to give robust feature representations for person retrieval. Human pose information could provide the locations of human skeleton to effectively guide the network to pay more attention on these key areas and could also help to reduce the noise distractions from background or occlusions. However, methods proposed by previous pose-related works might not be able to fully exploit the benefits of pose information and did not take into consideration the different contributions of different local features. In this paper, we propose a pose guided graph attention network, a multi-branch architecture consisting of one branch for global feature, one branch for mid-granular body features and one branch for fine-granular key point features. We use a pre-trained pose estimator to generate the key-point heatmap for local feature learning and carefully design a graph attention convolution layer to re-evaluate the contribution weights of extracted local features by modeling the similarities relations. Experiments results demonstrate the effectiveness of our approach on discriminative feature learning and we show that our model achieves state-of-the-art performances on several mainstream evaluation datasets. We also conduct a plenty of ablation studies and design different kinds of comparison experiments for our network to prove its effectiveness and robustness, including holistic datasets, partial datasets, occluded datasets and cross-domain tests.

AdaL: Adaptive Gradient Transformation Contributes to Convergences and Generalizations

Adaptive optimization methods have been widely used in deep learning. They scale the learning rates adaptively according to the past gradient, which has been shown to be effective to accelerate the convergence. However, they suffer from poor generalization performance compared with SGD. Recent studies point that smoothing exponential gradient noise leads to generalization degeneration phenomenon. Inspired by this, we propose AdaL, with a transformation on the original gradient. AdaL accelerates the convergence by amplifying the gradient in the early stage, as well as dampens the oscillation and stabilizes the optimization by shrinking the gradient later. Such modification alleviates the smoothness of gradient noise, which produces better generalization performance. We have theoretically proved the convergence of AdaL and demonstrated its effectiveness on several benchmarks.