Total: 400
Despite the advances in multi-person pose estimation, state-of-the-art techniques only deliver the human pose structure.Yet, they do not leverage the keypoints of human pose to deliver whole-body shape information for human instance segmentation. This paper presents PosePlusSeg, a joint model designed for both human pose estimation and instance segmentation. For pose estimation, PosePlusSeg first takes a bottom-up approach to detect the soft and hard keypoints of individuals by producing a strong keypoint heat map, then improves the keypoint detection confidence score by producing a body heat map. For instance segmentation, PosePlusSeg generates a mask offset where keypoint is defined as a centroid for the pixels in the embedding space, enabling instance-level segmentation for the human class. Finally, we propose a new pose and instance segmentation algorithm that enables PosePlusSeg to determine the joint structure of the human pose and instance segmentation. Experiments using the COCO challenging dataset demonstrate that PosePlusSeg copes better with challenging scenarios, like occlusions, en-tangled limbs, and overlapped people. PosePlusSeg outperforms state-of-the-art detection-based approaches achieving a 0.728 mAP for human pose estimation and a 0.445 mAP for instance segmentation. Code has been made available at: https://github.com/RaiseLab/PosePlusSeg.
With the increasing demands for understanding the internal behaviors of deep networks, Explainable AI (XAI) has been made remarkable progress in interpreting the model's decision. A family of attribution techniques has been proposed, highlighting whether the input pixels are responsible for the model's prediction. However, the existing attribution methods suffer from the lack of rule guidance and require further human interpretations. In this paper, we construct the 'if-then' logic rules that are sufficiently precise locally. Moreover, a novel rule-guided method, dynamic ablation (DA), is proposed to find a minimal bound sufficient in an input image to justify the network's prediction and aggregate iteratively to reach a complete attribution. Both qualitative and quantitative experiments are conducted to evaluate the proposed DA. We demonstrate the advantages of our method in providing clear and explicit explanations that are also easy for human experts to understand. Besides, through the attribution on a series of trained networks with different architectures, we show that more complex networks require less information to make a specific prediction.
We present Neural Marionette, an unsupervised approach that discovers the skeletal structure from a dynamic sequence and learns to generate diverse motions that are consistent with the observed motion dynamics. Given a video stream of point cloud observation of an articulated body under arbitrary motion, our approach discovers the unknown low-dimensional skeletal relationship that can effectively represent the movement. Then the discovered structure is utilized to encode the motion priors of dynamic sequences in a latent structure, which can be decoded to the relative joint rotations to represent the full skeletal motion. Our approach works without any prior knowledge of the underlying motion or skeletal structure, and we demonstrate that the discovered structure is even comparable to the hand-labeled ground truth skeleton in representing a 4D sequence of motion. The skeletal structure embeds the general semantics of possible motion space that can generate motions for diverse scenarios. We verify that the learned motion prior is generalizable to the multi-modal sequence generation, interpolation of two poses, and motion retargeting to a different skeletal structure.
In a convolutional object detector, the detection accuracy can be degraded often due to the low feature discriminability caused by geometric variation or transformation of an object. In this paper, we propose a deformable part region learning in order to allow decomposed part regions to be deformable according to geometric transformation of an object. To this end, we introduce trainable geometric parameters for the location of each part model. Because the ground truth of the part models is not available, we design classification and mask losses for part models, and learn the geometric parameters by minimizing an integral loss including those part losses. As a result, we can train a deformable part region network without extra super-vision and make each part model deformable according to object scale variation. Furthermore, for improving cascade object detection and instance segmentation, we present a Cascade deformable part region architecture which can refine whole and part detections iteratively in the cascade manner. Without bells and whistles, our implementation of a Cascade deformable part region detector achieves better detection and segmentation mAPs on COCO and VOC datasets, compared to the recent cascade and other state-of-the-art detectors.
We propose an end-to-end image compression and analysis model with Transformers, targeting to the cloud-based image classification application. Instead of placing an existing Transformer-based image classification model directly after an image codec, we aim to redesign the Vision Transformer (ViT) model to perform image classification from the compressed features and facilitate image compression with the long-term information from the Transformer. Specifically, we first replace the patchify stem (i.e., image splitting and embedding) of the ViT model with a lightweight image encoder modelled by a convolutional neural network. The compressed features generated by the image encoder are injected convolutional inductive bias and are fed to the Transformer for image classification bypassing image reconstruction. Meanwhile, we propose a feature aggregation module to fuse the compressed features with the selected intermediate features of the Transformer, and feed the aggregated features to a deconvolutional neural network for image reconstruction. The aggregated features can obtain the long-term information from the self-attention mechanism of the Transformer and improve the compression performance. The rate-distortion-accuracy optimization problem is finally solved by a two-step training strategy. Experimental results demonstrate the effectiveness of the proposed model in both the image compression and the classification tasks.
Handwritten mathematical expression recognition aims to automatically generate LaTeX sequences from given images. Currently, attention-based encoder-decoder models are widely used in this task. They typically generate target sequences in a left-to-right (L2R) manner, leaving the right-to-left (R2L) contexts unexploited. In this paper, we propose an Attention aggregation based Bi-directional Mutual learning Network (ABM) which consists of one shared encoder and two parallel inverse decoders (L2R and R2L). The two decoders are enhanced via mutual distillation, which involves one-to-one knowledge transfer at each training step, making full use of the complementary information from two inverse directions. Moreover, in order to deal with mathematical symbols in diverse scales, an Attention Aggregation Module (AAM) is proposed to effectively integrate multi-scale coverage attentions. Notably, in the inference phase, given that the model already learns knowledge from two inverse directions, we only use the L2R branch for inference, keeping the original parameter size and inference speed. Extensive experiments demonstrate that our proposed approach achieves the recognition accuracy of 56.85 % on CROHME 2014, 52.92 % on CROHME 2016, and 53.96 % on CROHME 2019 without data augmentation and model ensembling, substantially outperforming the state-of-the-art methods. The source code is available in https://github.com/XH-B/ABM.
Despite significant advancements of deep learning-based forgery detectors for distinguishing manipulated deepfake images, most detection approaches suffer from moderate to significant performance degradation with low-quality compressed deepfake images. Because of the limited information in low-quality images, detecting low-quality deepfake remains an important challenge. In this work, we apply frequency domain learning and optimal transport theory in knowledge distillation (KD) to specifically improve the detection of low-quality compressed deepfake images. We explore transfer learning capability in KD to enable a student network to learn discriminative features from low-quality images effectively. In particular, we propose the Attention-based Deepfake detection Distiller (ADD), which consists of two novel distillations: 1) frequency attention distillation that effectively retrieves the removed high-frequency components in the student network, and 2) multi-view attention distillation that creates multiple attention vectors by slicing the teacher’s and student’s tensors under different views to transfer the teacher tensor’s distribution to the student more efficiently. Our extensive experimental results demonstrate that our approach outperforms state-of-the-art baselines in detecting low-quality compressed deepfake images.
The predefined artificially-balanced training classes in object recognition have limited capability in modeling real-world scenarios where objects are imbalanced-distributed with unknown classes. In this paper, we discuss a promising solution to the Open-set Long-Tailed Recognition (OLTR) task utilizing metric learning. Firstly, we propose a distribution-sensitive loss, which weighs more on the tail classes to decrease the intra-class distance in the feature space. Building upon these concentrated feature clusters, a local-density-based metric is introduced, called Localizing Unfamiliarity Near Acquaintance (LUNA), to measure the novelty of a testing sample. LUNA is flexible with different cluster sizes and is reliable on the cluster boundary by considering neighbors of different properties. Moreover, contrary to most of the existing works that alleviate the open-set detection as a simple binary decision, LUNA is a quantitative measurement with interpretable meanings. Our proposed method exceeds the state-of-the-art algorithm by 4-6% in the closed-set recognition accuracy and 4% in F-measure under the open-set on the public benchmark datasets, including our own newly introduced fine-grained OLTR dataset about marine species (MS-LT), which is the first naturally-distributed OLTR dataset revealing the genuine genetic relationships of the classes.
We proposed a Prior Gradient Mask Guided Pruning-aware Fine-Tuning (PGMPF) framework to accelerate deep Convolutional Neural Networks (CNNs). In detail, the proposed PGMPF selectively suppresses the gradient of those ”unimportant” parameters via a prior gradient mask generated by the pruning criterion during fine-tuning. PGMPF has three charming characteristics over previous works: (1) Pruning-aware network fine-tuning. A typical pruning pipeline consists of training, pruning and fine-tuning, which are relatively independent, while PGMPF utilizes a variant of the pruning mask as a prior gradient mask to guide fine-tuning, without complicated pruning criteria. (2) An excellent tradeoff between large model capacity during fine-tuning and stable convergence speed to obtain the final compact model. Previous works preserve more training information of pruned parameters during fine-tuning to pursue better performance, which would incur catastrophic non-convergence of the pruned model for relatively large pruning rates, while our PGMPF greatly stabilizes the fine-tuning phase by gradually constraining the learning rate of those ”unimportant” parameters. (3) Channel-wise random dropout of the prior gradient mask to impose some gradient noise to fine-tuning to further improve the robustness of final compact model. Experimental results on three image classification benchmarks CIFAR10/ 100 and ILSVRC-2012 demonstrate the effectiveness of our method for various CNN architectures, datasets and pruning rates. Notably, on ILSVRC-2012, PGMPF reduces 53.5% FLOPs on ResNet-50 with only 0.90% top-1 accuracy drop and 0.52% top-5 accuracy drop, which has advanced the state-of-the-art with negligible extra computational cost.
Blackbox transfer attacks for image classifiers have been extensively studied in recent years. In contrast, little progress has been made on transfer attacks for object detectors. Object detectors take a holistic view of the image and the detection of one object (or lack thereof) often depends on other objects in the scene. This makes such detectors inherently context-aware and adversarial attacks in this space are more challenging than those targeting image classifiers. In this paper, we present a new approach to generate context-aware attacks for object detectors. We show that by using co-occurrence of objects and their relative locations and sizes as context information, we can successfully generate targeted mis-categorization attacks that achieve higher transfer success rates on blackbox object detectors than the state-of-the-art. We test our approach on a variety of object detectors with images from PASCAL VOC and MS COCO datasets and demonstrate up to 20 percentage points improvement in performance compared to the other state-of-the-art methods.
Recently, deep learning has been proven to be a promising approach in standard dynamic range (SDR) image compression. However, due to the wide luminance distribution of high dynamic range (HDR) images and the lack of large standard datasets, developing a deep model for HDR image compression is much more challenging. To tackle this issue, we view HDR data as distributional shifts of SDR data and the HDR image compression can be modeled as an out-of-distribution generalization (OoD) problem. Herein, we propose a novel out-of-distribution (OoD) HDR image compression framework (OoDHDR-codec). It learns the general representation across HDR and SDR environments, and allows the model to be trained effectively using a large set of SDR datases supplemented with much fewer HDR samples. Specifically, OoDHDR-codec consists of two branches to process the data from two environments. The SDR branch is a standard blackbox network. For the HDR branch, we develop a hybrid system that models luminance masking and tone mapping with white-box modules and performs content compression with black-box neural networks. To improve the generalization from SDR training data on HDR data, we introduce an invariance regularization term to learn the common representation for both SDR and HDR compression. Extensive experimental results show that the OoDHDR codec achieves strong competitive in-distribution performance and state-of-the-art OoD performance. To the best of our knowledge, our proposed approach is the first work to model HDR compression as OoD generalization problems and our OoD generalization algorithmic framework can be applied to any deep compression model in addition to the network architectural choice demonstrated in the paper. Code available at https://github.com/caolinfeng/OoDHDR-codec.
In this paper, we propose a novel method to mine the commonsense knowledge shared between the video and text modalities for video-text retrieval, namely visual consensus modeling. Different from the existing works, which learn the video and text representations and their complicated relationships solely based on the pairwise video-text data, we make the first attempt to model the visual consensus by mining the visual concepts from videos and exploiting their co-occurrence patterns within the video and text modalities with no reliance on any additional concept annotations. Specifically, we build a shareable and learnable graph as the visual consensus, where the nodes denoting the mined visual concepts and the edges connecting the nodes representing the co-occurrence relationships between the visual concepts. Extensive experimental results on the public benchmark datasets demonstrate that our proposed method, with the ability to effectively model the visual consensus, achieves state-of-the-art performances on the bidirectional video-text retrieval task. Our code is available at https://github.com/sqiangcao99/VCM.
Recently, deep learning techniques have been extensively studied for pansharpening, which aims to generate a high resolution multispectral (HRMS) image by fusing a low resolution multispectral (LRMS) image with a high resolution panchromatic (PAN) image. However, existing deep learning-based pansharpening methods directly learn the mapping from LRMS and PAN to HRMS. These network architectures always lack sufficient interpretability, which limits further performance improvements. To alleviate this issue, we propose a novel deep network for pansharpening by combining the model-based methodology with the deep learning method. Firstly, we build an observation model for pansharpening using the convolutional sparse coding (CSC) technique and design a proximal gradient algorithm to solve this model. Secondly, we unfold the iterative algorithm into a deep network, dubbed as Proximal PanNet, by learning the proximal operators using convolutional neural networks. Finally, all the learnable modules can be automatically learned in an end-to-end manner. Experimental results on some benchmark datasets show that our network performs better than other advanced methods both quantitatively and qualitatively.
The recently proposed DEtection TRansformer (DETR) achieves promising performance for end-to-end object detection. However, it has relatively lower detection performance on small objects and suffers from slow convergence. This paper observed that DETR performs surprisingly well even on small objects when measuring Average Precision (AP) at decreased Intersection-over-Union (IoU) thresholds. Motivated by this observation, we propose a simple way to improve DETR by refining the coarse features and predicted locations. Specifically, we propose a novel Coarse-to-Fine (CF) decoder layer constituted of a coarse layer and a carefully designed fine layer. Within each CF decoder layer, the extracted local information (region of interest feature) is introduced into the flow of global context information from the coarse layer to refine and enrich the object query features via the fine layer. In the fine layer, the multi-scale information can be fully explored and exploited via the Adaptive Scale Fusion(ASF) module and Local Cross-Attention (LCA) module. The multi-scale information can also be enhanced by another proposed Transformer Enhanced FPN (TEF) module to further improve the performance. With our proposed framework (named CF-DETR), the localization accuracy of objects (especially for small objects) can be largely improved. As a byproduct, the slow convergence issue of DETR can also be addressed. The effectiveness of CF-DETR is validated via extensive experiments on the coco benchmark. CF-DETR achieves state-of-the-art performance among end-to-end detectors, e.g., achieving 47.8 AP using ResNet-50 with 36 epochs in the standard 3x training schedule.
This paper starts by revealing a surprising finding: without any learning, a randomly initialized CNN can localize objects surprisingly well. That is, a CNN has an inductive bias to naturally focus on objects, named as Tobias ("The object is at sight") in this paper. This empirical inductive bias is further analyzed and successfully applied to self-supervised learning (SSL). A CNN is encouraged to learn representations that focus on the foreground object, by transforming every image into various versions with different backgrounds, where the foreground and background separation is guided by Tobias. Experimental results show that the proposed Tobias significantly improves downstream tasks, especially for object detection. This paper also shows that Tobias has consistent improvements on training sets of different sizes, and is more resilient to changes in image augmentations.
We propose a dual-domain generative model to estimate a texture map from a single image for colorizing a 3D human model. When estimating a texture map, a single image is insufficient as it reveals only one facet of a 3D object. To provide sufficient information for estimating a complete texture map, the proposed model simultaneously generates multi-view hallucinations in the image domain and an estimated texture map in the texture domain. During the generating process, each domain generator exchanges features to the other by a flow-based local attention mechanism. In this manner, the proposed model can estimate a texture map utilizing abundant multi-view image features from which multiview hallucinations are generated. As a result, the estimated texture map contains consistent colors and patterns over the entire region. Experiments show the superiority of our model for estimating a directly render-able texture map, which is applicable to 3D animation rendering. Furthermore, our model also improves an overall generation quality in the image domain for pose and viewpoint transfer tasks.
Scene Graph Generation (SGG) aims to build a structured representation of a scene using objects and pairwise relationships, which benefits downstream tasks. However, current SGG methods usually suffer from sub-optimal scene graph generation because of the long-tailed distribution of training data. To address this problem, we propose Resistance Training using Prior Bias (RTPB) for the scene graph generation. Specifically, RTPB uses a distributed-based prior bias to improve models' detecting ability on less frequent relationships during training, thus improving the model generalizability on tail categories. In addition, to further explore the contextual information of objects and relationships, we design a contextual encoding backbone network, termed as Dual Transformer (DTrans). We perform extensive experiments on a very popular benchmark, VG150, to demonstrate the effectiveness of our method for the unbiased scene graph generation. In specific, our RTPB achieves an improvement of over 10% under the mean recall when applied to current SGG methods. Furthermore, DTrans with RTPB outperforms nearly all state-of-the-art methods with a large margin. Code is available at https://github.com/ChCh1999/RTPB
Although point-based networks are demonstrated to be accurate for 3D point cloud modeling, they are still falling behind their voxel-based competitors in 3D detection. We observe that the prevailing set abstraction design for down-sampling points may maintain too much unimportant background information that can affect feature learning for detecting objects. To tackle this issue, we propose a novel set abstraction method named Semantics-Augmented Set Abstraction (SASA). Technically, we first add a binary segmentation module as the side output to help identify foreground points. Based on the estimated point-wise foreground scores, we then propose a semantics-guided point sampling algorithm to help retain more important foreground points during down-sampling. In practice, SASA shows to be effective in identifying valuable points related to foreground objects and improving feature learning for point-based 3D detection. Additionally, it is an easy-to-plug-in module and able to boost various point-based detectors, including single-stage and two-stage ones. Extensive experiments on the popular KITTI and nuScenes datasets validate the superiority of SASA, lifting point-based detection models to reach comparable performance to state-of-the-art voxel-based methods. Code is available at https://github.com/blakechen97/SASA.
Video anomaly detection aims to automatically identify unusual objects or behaviours by learning from normal videos. Previous methods tend to use simplistic reconstruction or prediction constraints, which leads to the insufficiency of learned representations for normal data. As such, we propose a novel bi-directional architecture with three consistency constraints to comprehensively regularize the prediction task from pixel-wise, cross-modal, and temporal-sequence levels. First, predictive consistency is proposed to consider the symmetry property of motion and appearance in forwards and backwards time, which ensures the highly realistic appearance and motion predictions at the pixel-wise level. Second, association consistency considers the relevance between different modalities and uses one modality to regularize the prediction of another one. Finally, temporal consistency utilizes the relationship of the video sequence and ensures that the predictive network generates temporally consistent frames. During inference, the pattern of abnormal frames is unpredictable and will therefore cause higher prediction errors. Experiments show that our method outperforms advanced anomaly detectors and achieves state-of-the-art results on UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets.
Video-based person re-identification~(re-ID) is an important technique in visual surveillance systems which aims to match video snippets of people captured by different cameras. Existing methods are mostly based on convolutional neural networks~(CNNs), whose building blocks either process local neighbor pixels at a time, or, when 3D convolutions are used to model temporal information, suffer from the misalignment problem caused by person movement. In this paper, we propose to overcome the limitations of normal convolutions with a human-oriented graph method. Specifically, features located at person joint keypoints are extracted and connected as a spatial-temporal graph. These keypoint features are then updated by message passing from their connected nodes with a graph convolutional network~(GCN). During training, the GCN can be attached to any CNN-based person re-ID model to assist representation learning on feature maps, whilst it can be dropped after training for better inference speed. Our method brings significant improvements over the CNN-based baseline model on the MARS dataset with generated person keypoints and a newly annotated dataset: PoseTrackReID. It also defines a new state-of-the-art method in terms of top-1 accuracy and mean average precision in comparison to prior works.
Temporal action detection aims to locate the boundaries of action in the video. The current method based on boundary matching enumerates and calculates all possible boundary matchings to generate proposals. However, these methods neglect the long-range context aggregation in boundary prediction. At the same time, due to the similar semantics of adjacent matchings, local semantic aggregation of densely-generated matchings cannot improve semantic richness and discrimination. In this paper, we propose the end-to-end proposal generation method named Dual Context Aggregation Network (DCAN) to aggregate context on two levels, namely, boundary level and proposal level, for generating high-quality action proposals, thereby improving the performance of temporal action detection. Specifically, we design the Multi-Path Temporal Context Aggregation (MTCA) to achieve smooth context aggregation on boundary level and precise evaluation of boundaries. For matching evaluation, Coarse-to-fine Matching (CFM) is designed to aggregate context on the proposal level and refine the matching map from coarse to fine. We conduct extensive experiments on ActivityNet v1.3 and THUMOS-14. DCAN obtains an average mAP of 35.39% on ActivityNet v1.3 and reaches mAP 54.14% at IoU@0.5 on THUMOS-14, which demonstrates DCAN can generate high-quality proposals and achieve state-of-the-art performance. We release the code at https://github.com/cg1177/DCAN.
We present a customized 3D mesh Transformer model for the pose transfer task. As the 3D pose transfer essentially is a deformation procedure dependent on the given meshes, the intuition of this work is to perceive the geometric inconsistency between the given meshes with the powerful self-attention mechanism. Specifically, we propose a novel geometry-contrastive Transformer that has an efficient 3D structured perceiving ability to the global geometric inconsistencies across the given meshes. Moreover, locally, a simple yet efficient central geodesic contrastive loss is further proposed to improve the regional geometric-inconsistency learning. At last, we present a latent isometric regularization module together with a novel semi-synthesized dataset for the cross-dataset 3D pose transfer task towards unknown spaces. The massive experimental results prove the efficacy of our approach by showing state-of-the-art quantitative performances on SMPL-NPT, FAUST and our new proposed dataset SMG-3D datasets, as well as promising qualitative results on MG-cloth and SMAL datasets. It's demonstrated that our method can achieve robust 3D pose transfer and be generalized to challenging meshes from unknown spaces on cross-dataset tasks. The code and dataset are made available. Code is available: https://github.com/mikecheninoulu/CGT.
Weakly supervised temporal sentence grounding aims to temporally localize the target segment corresponding to a given natural language query, where it provides video-query pairs without temporal annotations during training. Most existing methods use the fused visual-linguistic feature to reconstruct the query, where the least reconstruction error determines the target segment. This work introduces a novel approach that explores the inter-contrast between videos in a composed video by selecting components from two different videos and fusing them into a single video. Such a straightforward yet effective composition strategy provides the temporal annotations at multiple composed positions, resulting in numerous videos with temporal ground-truths for training the temporal sentence grounding task. A transformer framework is introduced with multi-tasks training to learn a compact but efficient visual-linguistic space. The experimental results on the public Charades-STA and ActivityNet-Caption dataset demonstrate the effectiveness of the proposed method, where our approach achieves comparable performance over the state-of-the-art weakly-supervised baselines. The code is available at https://github.com/PPjmchen/Composition_WSTG.
Scene graph in a video conveys a wealth of information about objects and their relationships in the scene, thus benefiting many downstream tasks such as video captioning and visual question answering. Existing methods of scene graph generation require large-scale training videos annotated with objects and relationships in each frame to learn a powerful model. However, such comprehensive annotation is time-consuming and labor-intensive. On the other hand, it is much easier and less cost to annotate images with scene graphs, so we investigate leveraging annotated images to facilitate training a scene graph generation model for unannotated videos, namely image-to-video scene graph generation. This task presents two challenges: 1) infer unseen dynamic relationships in videos from static relationships in images due to the absence of motion information in images; 2) adapt objects and static relationships from images to video frames due to the domain shift between them. To address the first challenge, we exploit external commonsense knowledge to infer the unseen dynamic relationship from the temporal evolution of static relationships. We tackle the second challenge by hierarchical adversarial learning to reduce the data distribution discrepancy between images and video frames. Extensive experiment results on two benchmark video datasets demonstrate the effectiveness of our method.
In the last decade, the blossom of deep learning has witnessed the rapid development of scene text recognition. However, the recognition of low-resolution scene text images remains a challenge. Even though some super-resolution methods have been proposed to tackle this problem, they usually treat text images as general images while ignoring the fact that the visual quality of strokes (the atomic unit of text) plays an essential role for text recognition. According to Gestalt Psychology, humans are capable of composing parts of details into the most similar objects guided by prior knowledge. Likewise, when humans observe a low-resolution text image, they will inherently use partial stroke-level details to recover the appearance of holistic characters. Inspired by Gestalt Psychology, we put forward a Stroke-Aware Scene Text Image Super-Resolution method containing a Stroke-Focused Module (SFM) to concentrate on stroke-level internal structures of characters in text images. Specifically, we attempt to design rules for decomposing English characters and digits at stroke-level, then pre-train a text recognizer to provide stroke-level attention maps as positional clues with the purpose of controlling the consistency between the generated super-resolution image and high-resolution ground truth. The extensive experimental results validate that the proposed method can indeed generate more distinguishable images on TextZoom and manually constructed Chinese character dataset Degraded-IC13. Furthermore, since the proposed SFM is only used to provide stroke-level guidance when training, it will not bring any time overhead during the test phase. Code is available at https://github.com/FudanVI/FudanOCR/tree/main/text-gestalt.