Referring expression comprehension (REC) aims to locate a certain object in an image referred by a natural language expression. For joint understanding of regions and expressions, existing REC works typically target on modeling the cross-modal relevance in each region-expression pair within each single image. In this paper, we explore a new but general REC-related problem, named Group-based REC, where the regions and expressions can come from different subject-related images (images in the same group), e.g., sets of photo albums or video frames. Different from REC, Group-based REC involves differentiated cross-modal relevances within each group and across different groups, which, however, are neglected in the existing one-line paradigm. To this end, we propose a novel relevance-guided multi-group self-paced learning schema (termed RMSL), where the within-group region-expression pairs are adaptively assigned with different priorities according to their cross-modal relevances, and the bias of the group priority is balanced via an across-group relevance constraint simultaneously. In particular, based on the visual and textual semantic features, RMSL conducts an adaptive learning cycle upon triplet ranking, where (1) the target-negative region-expression pairs with low within-group relevances are used preferentially in model training to distinguish the primary semantics of the target objects, and (2) an across-group relevance regularization is integrated into model training to balance the bias of group priority. The relevances, the pairs, and the model parameters are alternatively updated upon a unified self-paced hinge loss.
Light-weight super-resolution (SR) models have received considerable attention for their serviceability in mobile devices. Many efforts employ network quantization to compress SR models. However, these methods suffer from severe performance degradation when quantizing the SR models to ultra-low precision (e.g., 2-bit and 3-bit) with the low-cost layer-wise quantizer. In this paper, we identify that the performance drop comes from the contradiction between the layer-wise symmetric quantizer and the highly asymmetric activation distribution in SR models. This discrepancy leads to either a waste on the quantization levels or detail loss in reconstructed images. Therefore, we propose a novel activation quantizer, referred to as Dynamic Dual Trainable Bounds (DDTB), to accommodate the asymmetry of the activations. Specifically, DDTB innovates in: 1) A layer-wise quantizer with trainable upper and lower bounds to tackle the highly asymmetric activations. 2) A dynamic gate controller to adaptively adjust the upper and lower bounds at runtime to overcome the drastically varying activation ranges over different samples.To reduce the extra overhead, the dynamic gate controller is quantized to 2-bit and applied to only part of the SR networks according to the introduced dynamic intensity. Extensive experiments demonstrate that our DDTB exhibits significant performance improvements in ultra-low precision. For example, our DDTB achieves a 0.70dB PSNR increase on Urban100 benchmark when quantizing EDSR to 2-bit and scaling up output images to x4. Code is at \url{https://github.com/zysxmu/DDTB}.
Vision Transformers (ViT) have made many breakthroughs in computer vision tasks. However, considerable redundancy arises in the spatial dimension of an input image, leading to massive computational costs. Therefore, We propose a coarse-to-fine vision transformer (CF-ViT) to relieve computational burden while retaining performance in this paper. Our proposed CF-ViT is motivated by two important observations in modern ViT models: (1) The coarse-grained patch splitting can locate informative regions of an input image. (2) Most images can be well recognized by a ViT model in a small-length token sequence. Therefore, our CF-ViT implements network inference in a two-stage manner. At coarse inference stage, an input image is split into a small-length patch sequence for a computationally economical classification. If not well recognized, the informative patches are identified and further re-split in a fine-grained granularity. Extensive experiments demonstrate the efficacy of our CF-ViT. For example, without any compromise on performance, CF-ViT reduces 53% FLOPs of LV-ViT, and also achieves 2.01x throughput.
This paper focuses on the challenging crowd counting task. As large-scale variations often exist within crowd images, neither fixed-size convolution kernel of CNN nor fixed-size attention of recent vision transformers can well handle this kind of variation. To address this problem, we propose a Multifaceted Attention Network (MAN) to improve transformer models in local spatial relation encoding. MAN incorporates global attention from a vanilla transformer, learnable local attention, and instance attention into a counting model. Firstly, the local Learnable Region Attention (LRA) is proposed to assign attention exclusively for each feature location dynamically. Secondly, we design the Local Attention Regularization to supervise the training of LRA by minimizing the deviation among the attention for different feature locations. Finally, we provide an Instance Attention mechanism to focus on the most important instances dynamically during training. Extensive experiments on four challenging crowd counting datasets namely ShanghaiTech, UCF-QNRF, JHU++, and NWPU have validated the proposed method. Codes: https://github.com/LoraLinH/Boosting-Crowd-Counting-via-Multifaceted-Attention.
This paper focuses on filter-level network pruning. A novel pruning method, termed CLR-RNF, is proposed. We first reveal a "long-tail" long-tail pruning problem in magnitude-based weight pruning methods, and then propose a computation-aware measurement for individual weight importance, followed by a Cross-Layer Ranking (CLR) of weights to identify and remove the bottom-ranked weights. Consequently, the per-layer sparsity makes up of the pruned network structure in our filter pruning. Then, we introduce a recommendation-based filter selection scheme where each filter recommends a group of its closest filters. To pick the preserved filters from these recommended groups, we further devise a k-Reciprocal Nearest Filter (RNF) selection scheme where the selected filters fall into the intersection of these recommended groups. Both our pruned network structure and the filter selection are non-learning processes, which thus significantly reduce the pruning complexity, and differentiate our method from existing works. We conduct image classification on CIFAR-10 and ImageNet to demonstrate the superiority of our CLR-RNF over the state-of-the-arts. For example, on CIFAR-10, CLR-RNF removes 74.1% FLOPs and 95.0% parameters from VGGNet-16 with even 0.3\% accuracy improvements. On ImageNet, it removes 70.2% FLOPs and 64.8% parameters from ResNet-50 with only 1.7% top-5 accuracy drops. Our project is at https://github.com/lmbxmu/CLR-RNF.
Network sparsity receives popularity mostly due to its capability to reduce the network complexity. Extensive studies excavate gradient-driven sparsity. Typically, these methods are constructed upon premise of weight independence, which however, is contrary to the fact that weights are mutually influenced. Thus, their performance remains to be improved. In this paper, we propose to further optimize gradient-driven sparsity (OptG) by solving this independence paradox. Our motive comes from the recent advances on supermask training which shows that sparse subnetworks can be located in a randomly initialized network by simply updating mask values without modifying any weight. We prove that supermask training is to accumulate the weight gradients and can partly solve the independence paradox. Consequently, OptG integrates supermask training into gradient-driven sparsity, and a specialized mask optimizer is designed to solve the independence paradox. Experiments show that OptG can well surpass many existing state-of-the-art competitors. Our code is available at \url{https://github.com/zyxxmu/OptG}.
This paper reports the results and post-challenge analyses of ChaLearn's AutoDL challenge series, which helped sorting out a profusion of AutoML solutions for Deep Learning (DL) that had been introduced in a variety of settings, but lacked fair comparisons. All input data modalities (time series, images, videos, text, tabular) were formatted as tensors and all tasks were multi-label classification problems. Code submissions were executed on hidden tasks, with limited time and computational resources, pushing solutions that get results quickly. In this setting, DL methods dominated, though popular Neural Architecture Search (NAS) was impractical. Solutions relied on fine-tuned pre-trained networks, with architectures matching data modality. Post-challenge tests did not reveal improvements beyond the imposed time limit. While no component is particularly original or novel, a high level modular organization emerged featuring a "meta-learner", "data ingestor", "model selector", "model/learner", and "evaluator". This modularity enabled ablation studies, which revealed the importance of (off-platform) meta-learning, ensembling, and efficient data management. Experiments on heterogeneous module combinations further confirm the (local) optimality of the winning solutions. Our challenge legacy includes an ever-lasting benchmark (http://autodl.chalearn.org), the open-sourced code of the winners, and a free "AutoDL self-service".
In this paper, we discover two factors that inhibit POMs from achieving high perceptual quality: 1) center-oriented optimization (COO) problem and 2) model's low-frequency tendency. First, POMs tend to generate an SR image whose position in the feature space is closest to the distribution center of all potential high-resolution (HR) images, resulting in such POMs losing high-frequency details. Second, $90\%$ area of an image consists of low-frequency signals; in contrast, human perception relies on an image's high-frequency details. However, POMs apply the same calculation to process different-frequency areas, so that POMs tend to restore the low-frequency regions. Based on these two factors, we propose a Detail Enhanced Contrastive Loss (DECLoss), by combining a high-frequency enhancement module and spatial contrastive learning module, to reduce the influence of the COO problem and low-Frequency tendency. Experimental results show the efficiency and effectiveness when applying DECLoss on several regular SR models. E.g, in EDSR, our proposed method achieves 3.60$\times$ faster learning speed compared to a GAN-based method with a subtle degradation in visual quality. In addition, our final results show that an SR network equipped with our DECLoss generates more realistic and visually pleasing textures compared to state-of-the-art methods. %The source code of the proposed method is included in the supplementary material and will be made publicly available in the future.
With various facial manipulation techniques arising, face forgery detection has drawn growing attention due to security concerns. Previous works always formulate face forgery detection as a classification problem based on cross-entropy loss, which emphasizes category-level differences rather than the essential discrepancies between real and fake faces, limiting model generalization in unseen domains. To address this issue, we propose a novel face forgery detection framework, named Dual Contrastive Learning (DCL), which specially constructs positive and negative paired data and performs designed contrastive learning at different granularities to learn generalized feature representation. Concretely, combined with the hard sample selection strategy, Inter-Instance Contrastive Learning (Inter-ICL) is first proposed to promote task-related discriminative features learning by especially constructing instance pairs. Moreover, to further explore the essential discrepancies, Intra-Instance Contrastive Learning (Intra-ICL) is introduced to focus on the local content inconsistencies prevalent in the forged faces by constructing local-region pairs inside instances. Extensive experiments and visualizations on several datasets demonstrate the generalization of our method against the state-of-the-art competitors.
Transfer adversarial attack is a non-trivial black-box adversarial attack that aims to craft adversarial perturbations on the surrogate model and then apply such perturbations to the victim model. However, the transferability of perturbations from existing methods is still limited, since the adversarial perturbations are easily overfitting with a single surrogate model and specific data pattern. In this paper, we propose a Learning to Learn Transferable Attack (LLTA) method, which makes the adversarial perturbations more generalized via learning from both data and model augmentation. For data augmentation, we adopt simple random resizing and padding. For model augmentation, we randomly alter the back propagation instead of the forward propagation to eliminate the effect on the model prediction. By treating the attack of both specific data and a modified model as a task, we expect the adversarial perturbations to adopt enough tasks for generalization. To this end, the meta-learning algorithm is further introduced during the iteration of perturbation generation. Empirical results on the widely-used dataset demonstrate the effectiveness of our attack method with a 12.85% higher success rate of transfer attack compared with the state-of-the-art methods. We also evaluate our method on the real-world online system, i.e., Google Cloud Vision API, to further show the practical potentials of our method.