Dense Prediction Research Articles

A Survey on Continual Semantic Segmentation: Theory, Challenge, Method and Application.

Continual learning, also known as incremental learning or life-long learning, stands at the forefront of deep learning and AI systems. It breaks through the obstacle of one-way training on close sets and enables continuous adaptive learning on open-set conditions. In the recent decade, continual learning has been explored and applied in multiple fields especially in computer vision covering classification, detection and segmentation tasks. Continual semantic segmentation (CSS), of which the dense prediction peculiarity makes it a challenging, intricate and burgeoning task. In this paper, we present a review of CSS, committing to building a comprehensive survey on problem formulations, primary challenges, universal datasets, neoteric theories and multifarious applications. Concretely, we begin by elucidating the problem definitions and primary challenges. Based on an in-depth investigation of relevant approaches, we sort out and categorize current CSS models into two main branches including data-replay and data-free sets. In each branch, the corresponding approaches are similarity-based clustered and thoroughly analyzed, following qualitative comparison and quantitative reproductions on relevant datasets. Besides, we also introduce four CSS specialities with diverse application scenarios and development tendencies. Furthermore, we develop a benchmark for CSS encompassing representative references, evaluation results and reproductions, which is available athttps://github.com/YBIO/SurveyCSS. We hope this survey can serve as a reference-worthy and stimulating contribution to the advancement of the life-long learning field, while also providing valuable perspectives for related fields.

SOccDPT: 3D Semantic Occupancy From Dense Prediction Transformers Trained Under Memory Constraints

We present SOccDPT, a memory-efficient approach for 3D semantic occupancy prediction from monocular image input using dense prediction transformers. To address the limitations of existing methods trained on structured traffic datasets, we train our model on unstructured datasets including the Indian Driving Dataset and Bengaluru Driving Dataset. Our semisupervised training pipeline allows SOccDPT to learn from datasets with limited labels by reducing the requirement for manual labeling by substituting it with pseudo-ground truth labels to produce our Bengaluru Semantic Occupancy Dataset. This broader training enhances our model’s ability to handle unstructured traffic scenarios effectively. To overcome memory limitations during training, we introduce patch-wise training where we select a subset of parameters to train each epoch, reducing memory usage during auto-grad graph construction. In the context of unstructured traffic and memory-constrained training and inference, SOccDPT outperforms existing disparity estimation approaches as shown by the RMSE score of 9.1473, achieves a semantic segmentation IoU score of 46.02% and operates at a competitive frequency of 69.47 Hz. We make our code and semantic occupancy dataset public.

Point-to-Pixel Prompting for Point Cloud Analysis With Pre-Trained Image Models.

Nowadays, pre-training big models on large-scale datasets has achieved great success and dominated many downstream tasks in natural language processing and 2D vision, while pre-training in 3D vision is still under development. In this paper, we provide a new perspective of transferring the pre-trained knowledge from 2D domain to 3D domain with Point-to-Pixel Prompting in data space and Pixel-to-Point distillation in feature space, exploiting shared knowledge in images and point clouds that display the same visual world. Following the principle of prompting engineering, Point-to-Pixel Prompting transforms point clouds into colorful images with geometry-preserved projection and geometry-aware coloring. Then the pre-trained image models can be directly implemented for point cloud tasks without structural changes or weight modifications. With projection correspondence in feature space, Pixel-to-Point distillation further regards pre-trained image models as the teacher model and distills pre-trained 2D knowledge to student point cloud models, remarkably enhancing inference efficiency and model capacity for point cloud analysis. We conduct extensive experiments in both object classification and scene segmentation under various settings to demonstrate the superiority of our method. In object classification, we reveal the important scale-up trend of Point-to-Pixel Prompting and attain 90.3% accuracy on ScanObjectNN dataset, surpassing previous literature by a large margin. In scene-level semantic segmentation, our method outperforms traditional 3D analysis approaches and shows competitive capacity in dense prediction tasks.

DrasCLR: A self-supervised framework of learning disease-related and anatomy-specific representation for 3D lung CT images

Large-scale volumetric medical images with annotation are rare, costly, and time prohibitive to acquire. Self-supervised learning (SSL) offers a promising pre-training and feature extraction solution for many downstream tasks, as it only uses unlabeled data. Recently, SSL methods based on instance discrimination have gained popularity in the medical imaging domain. However, SSL pre-trained encoders may use many clues in the image to discriminate an instance that are not necessarily disease-related. Moreover, pathological patterns are often subtle and heterogeneous, requiring the ability of the desired method to represent anatomy-specific features that are sensitive to abnormal changes in different body parts. In this work, we present a novel SSL framework, named DrasCLR, for 3D lung CT images to overcome these challenges. We propose two domain-specific contrastive learning strategies: one aims to capture subtle disease patterns inside a local anatomical region, and the other aims to represent severe disease patterns that span larger regions. We formulate the encoder using conditional hyper-parameterized network, in which the parameters are dependant on the anatomical location, to extract anatomically sensitive features. Extensive experiments on large-scale datasets of lung CT scans show that our method improves the performance of many downstream prediction and segmentation tasks. The patient-level representation improves the performance of the patient survival prediction task. We show how our method can detect emphysema subtypes via dense prediction. We demonstrate that fine-tuning the pre-trained model can significantly reduce annotation efforts without sacrificing emphysema detection accuracy. Our ablation study highlights the importance of incorporating anatomical context into the SSL framework. Our codes are available at https://github.com/batmanlab/DrasCLR.

Self-supervised representation learning using feature pyramid siamese networks for colorectal polyp detection

Colorectal cancer is a leading cause of cancer-related deaths globally. In recent years, the use of convolutional neural networks in computer-aided diagnosis (CAD) has facilitated simpler detection of early lesions like polyps during real-time colonoscopy. However, the majority of existing techniques require a large training dataset annotated by experienced experts. To alleviate the laborious task of image annotation and utilize the vast amounts of readily available unlabeled colonoscopy data to further improve the polyp detection ability, this study proposed a novel self-supervised representation learning method called feature pyramid siamese networks (FPSiam). First, a feature pyramid encoder module was proposed to effectively extract and fuse both local and global feature representations among colonoscopic images, which is important for dense prediction tasks like polyp detection. Next, a self-supervised visual feature representation containing the general feature of colonoscopic images is learned by the siamese networks. Finally, the feature representation will be transferred to the downstream colorectal polyp detection task. A total of 103 videos (861,400 frames), 100 videos (24,789 frames), and 60 videos (15,397 frames) in the LDPolypVideo dataset are used to pre-train, train, and test the performance of the proposed FPSiam and its counterparts, respectively. The experimental results have illustrated that our FPSiam approach obtains the optimal capability, which is better than that of other state-of-the-art self-supervised learning methods and is also higher than the method based on transfer learning by 2.3 mAP and 3.6 mAP for two typical detectors. In conclusion, FPSiam provides a cost-efficient solution for developing colorectal polyp detection systems, especially in conditions where only a small fraction of the dataset is labeled while the majority remains unlabeled. Besides, it also brings fresh perspectives into other endoscopic image analysis tasks.

HCS-Net: Multi-level deformation strategy combined with quadruple attention for image registration

Background and objectiveNon-rigid image registration plays a significant role in computer-aided diagnosis and surgical navigation for brain diseases. Registration methods that utilize convolutional neural networks (CNNs) have shown excellent accuracy when applied to brain magnetic resonance images (MRI). However, CNNs have limitations in understanding long-range spatial relationships in images, which makes it challenging to incorporate contextual information. And in intricate image registration tasks, it is difficult to achieve a satisfactory dense prediction field, resulting in poor registration performance. MethodsThis paper proposes a multi-level deformable unsupervised registration model that combines Transformer and CNN to achieve non-rigid registration of brain MRI. Firstly, utilizing a dual encoder structure to establish the dependency relationship between the global features of two images and to merge features of varying scales, as well as to preserve the relative spatial position information of feature maps at different scales. Then the proposed multi-level deformation strategy utilizes different deformable fields of varying resolutions generated by the decoding structure to progressively deform the moving image. Ultimately, the proposed quadruple attention module is incorporated into the decoding structure to merge feature information from various directions and emphasize the spatial features in the dominant channels. ResultsThe experimental results on multiple brain MR datasets demonstrate that the promising network could provide accurate registration and is comparable to state-of-the-art methods. ConclusionThe proposed registration model can generate superior deformable fields and achieve more precise registration effects, enhancing the auxiliary role of medical image registration in various fields and advancing the development of computer-aided diagnosis, surgical navigation, and related domains.

Dynamic context modeling based lightweight high-resolution network for dense prediction

Recent research shows that high-resolution networks can provide representative multi-scale features for vision tasks. However, high-resolution-based architectures are weak in capturing spatial long-range information, and they are computationally expensive. To alleviate these problems and encourage the engineering applications of dense prediction, based on vanilla high-resolution network, this paper presents a novel efficient architecture for dense prediction, namely Dynamic context modeling based lightweight High-Resolution Network (Dynamic-HRNet). In particular, the network consists of two key components: (i) dynamic split convolution, a more efficient and flexible convolution compared to the standard convolution, which could be conveniently embedded into other networks, and (ii) adaptive context modeling, which adaptively captures local and global contextual information to enrich the representation in parallel. Using the above two components, two lightweight blocks are designed that serve as the basic building units of Dynamic-HRNet, termed dynamic multi-scale context block and dynamic global context block, respectively. Extensive experiments demonstrate that the proposed Dynamic-HRNet, as a backbone, achieves significant superiority in popular benchmarks for human pose estimation (70.6% AP on COCO dataset and 87.6% PCKh on MPII dataset) and semantic segmentation (75.3% MIoU on Cityscapes dataset). Considering both efficiency and accuracy, it outperforms the most advanced lightweight architectures.

CycleMLP: A MLP-like Architecture for Dense Visual Predictions.

This paper presents a simple yet effective multilayer perceptron (MLP) architecture, namely CycleMLP, which is a versatile neural backbone network capable of solving various tasks of dense visual predictions such as object detection, segmentation, and human pose estimation. Compared to recent advanced MLP architectures such as MLP-Mixer [89], ResMLP [90], and gMLP [58], whose architectures are sensitive to image size and are infeasible in dense prediction tasks, CycleMLP has two appealing advantages. (1) CycleMLP can cope with various spatial sizes of images. (2) CycleMLP achieves linear computational complexity with respect to the image size by using local windows. In contrast, previous MLPs have O(N2) computational complexity due to their full connections in space. (3) The relationship between convolution, multi-head self-attention in Transformer, and CycleMLP are discussed through an intuitive theoretical analysis. We build a family of models that can surpass state-of-the-art MLP and Transformer models e.g., Swin Transformer [60], while using fewer parameters and FLOPs. CycleMLP expands the MLP-like models' applicability, making them versatile backbone networks that achieve competitive results on dense prediction tasks For example, CycleMLP-Tiny outperforms Swin-Tiny by 1.3% mIoU on ADE20K dataset with fewer FLOPs. Moreover, CycleMLP also shows excellent zero-shot robustness on ImageNet-C dataset. The source codes and models are available at https://github.com/ShoufaChen/CycleMLP.

Universal Representations: A Unified Look at Multiple Task and Domain Learning

AbstractWe propose a unified look at jointly learning multiple vision tasks and visual domains through universal representations, a single deep neural network. Learning multiple problems simultaneously involves minimizing a weighted sum of multiple loss functions with different magnitudes and characteristics and thus results in unbalanced state of one loss dominating the optimization and poor results compared to learning a separate model for each problem. To this end, we propose distilling knowledge of multiple task/domain-specific networks into a single deep neural network after aligning its representations with the task/domain-specific ones through small capacity adapters. We rigorously show that universal representations achieve state-of-the-art performances in learning of multiple dense prediction problems in NYU-v2 and Cityscapes, multiple image classification problems from diverse domains in Visual Decathlon Dataset and cross-domain few-shot learning in MetaDataset. Finally we also conduct multiple analysis through ablation and qualitative studies.

Very high resolution canopy height maps from RGB imagery using self-supervised vision transformer and convolutional decoder trained on aerial lidar

Vegetation structure mapping is critical for understanding the global carbon cycle and monitoring nature-based approaches to climate adaptation and mitigation. Repeated measurements of these data allow for the observation of deforestation or degradation of existing forests, natural forest regeneration, and the implementation of sustainable agricultural practices like agroforestry. Assessments of tree canopy height and crown projected area at a high spatial resolution are also important for monitoring carbon fluxes and assessing tree-based land uses, since forest structures can be highly spatially heterogeneous, especially in agroforestry systems. Very high resolution satellite imagery (less than one meter (1 m) Ground Sample Distance) makes it possible to extract information at the tree level while allowing monitoring at a very large scale. This paper presents the first high-resolution canopy height map concurrently produced for multiple sub-national jurisdictions. Specifically, we produce very high resolution canopy height maps for the states of California and São Paulo, a significant improvement in resolution over the ten meter (10 m) resolution of previous Sentinel / GEDI based worldwide maps of canopy height. The maps are generated by the extraction of features from a self-supervised model trained on Maxar imagery from 2017 to 2020, and the training of a dense prediction decoder against aerial lidar maps. We also introduce a post-processing step using a convolutional network trained on GEDI observations. We evaluate the proposed maps with set-aside validation lidar data as well as by comparing with other remotely sensed maps and field-collected data, and find our model produces an average Mean Absolute Error (MAE) of 2.8 m and Mean Error (ME) of 0.6 m.

Box-supervised dynamical instance segmentation for in-field cotton

Crop growing information is important for precision agriculture. Instance segmentation can be employed to extract not only the location and quantity information of objects from images, but also the individual phenotype information of each target. However, when it is applied on real in-field images, complex environment increases the cost of annotating. Besides, it is hard for instance segmentation network with complicated structure to be deployed in real application scenarios. To cost-efficiently extract data used for precision management from field images, this paper proposes a novel model based on dense prediction and conditionally parameterized convolution termed Box-supervised dynamical regression instance segmentation net(BDRISNet) for segmenting individuals from image in a simple way, using box-level supervision. By formulating segmentation into dense regression from each pixel to the location of its instance indicators, mask features can be generated without redundant convolution branch. The compact construction of proposed network achieves efficient inference for segmentation. Robustness and accuracy of proposed method are evaluated on an in-field cotton boll image dataset we provide. Comparison experiments with other state-of-the-arts demonstrate that our method has an average improvement of 3.13%, 5.55%, 7.08% in AP, AP50 and AP75. Furthermore, predictions of proposed method have desirable and balanced recall and precision.

Knowledge distillation of multi-scale dense prediction transformer for self-supervised depth estimation

Depth estimation is an inverse projection problem that estimates pixel-level distances from a single image. Although, supervised methods have shown promising results, it has intrinsic limitations in requiring ground truth depth from an external sensor. On the other hand, self-supervised depth estimation relieves the burden for collecting calibrated training data, while there is still a large performance gap between supervised and self-supervised methods. The objective of this study is to reduce the performance gap between the supervised and self-supervised approaches. The loss function of previous self-supervised methods is mainly based on a photometric error, which is indirectly computed from synthesized images using depth and pose estimates. In this paper, we argue that direct depth cue is more effective to train a depth estimation network. To obtain the direct depth cue, we employed a knowledge distillation technique, which is a teacher-student learning framework. The teacher network was trained in a self-supervised manner based on a photometric error, and its predictions were utilized to train a student network. We constructed a multi-scale dense prediction transformer with Monte Carlo dropout, and multi-scale distillation loss was proposed to train the student network based on the ensemble of stochastic estimates. Experiments were conducted on the KITTI and Make3D datasets, and our proposed method achieved the state-of-the-art accuracy in self-supervised depth estimation. Our code is publicly available at https://github.com/ji-min-song/KD-of-MS-DPT.

Joint representation learning for text and 3D point cloud

Recent advancements in vision-language pre-training (e.g., CLIP) have enabled 2D vision models to benefit from language supervision. However, the joint representation learning of 3D point cloud with text remains under-explored due to challenges in acquiring 3D-Text data pairs. Prior works propose to project point clouds into 2D depth maps and directly use CLIP, while they sacrifice 3D structural information, limiting its applicability. In this paper, we put forward Text4Point, a novel framework to construct language-guided 3D models for dense prediction tasks. Text4Point utilizes 2D images as a bridge to connect the point cloud and language modalities. It follows a pre-training and fine-tuning paradigm. During pre-training, we leverage dense contrastive learning to align the image and point cloud representations using the readily available RGB-D data. Together with the well-aligned image and text features achieved by CLIP, the point cloud features are implicitly aligned with the text embeddings. Further, we propose a Text Querying Module to integrate language information into 3D representation learning by querying text embeddings with point cloud features. For fine-tuning, the model learns 3D representations under informative language guidance without 2D images. Extensive experiments demonstrate consistent improvement on various dense prediction tasks with Text4Point.

Deep intra-image contrastive learning for weakly supervised one-step person search

Weakly supervised person search aims to perform joint pedestrian detection and re-identification (re-id) with only bounding-box annotations. Recently, the idea of contrastive learning is initially applied to weakly supervised person search, where two common contrast strategies are memory-based contrast and intra-image contrast. We argue that current intra-image contrast is shallow, which suffers from spatial-level and occlusion-level variance. In this paper, we present a novel deep intra-image contrastive learning using a Siamese network. Two key modules are spatial-invariant contrast (SIC) and occlusion-invariant contrast (OIC). SIC performs many-to-one contrasts between two branches of Siamese network and dense prediction contrasts in one branch of Siamese network. With these many-to-one and dense contrasts, SIC tends to learn discriminative scale-invariant and location-invariant features to solve spatial-level variance. OIC enhances feature consistency with the masking strategy to learn occlusion-invariant features. Extensive experiments are performed on two person search datasets. Our method achieves a state-of-the-art performance among weakly supervised one-step person search approaches.

UniFormer: Unifying Convolution and Self-Attention for Visual Recognition.

It is a challenging task to learn discriminative representation from images and videos, due to large local redundancy and complex global dependency in these visual data. Convolution neural networks (CNNs) and vision transformers (ViTs) have been two dominant frameworks in the past few years. Though CNNs can efficiently decrease local redundancy by convolution within a small neighborhood, the limited receptive field makes it hard to capture global dependency. Alternatively, ViTs can effectively capture long-range dependency via self-attention, while blind similarity comparisons among all the tokens lead to high redundancy. To resolve these problems, we propose a novel Unified transFormer (UniFormer), which can seamlessly integrate the merits of convolution and self-attention in a concise transformer format. Different from the typical transformer blocks, the relation aggregators in our UniFormer block are equipped with local and global token affinity respectively in shallow and deep layers, allowing tackling both redundancy and dependency for efficient and effective representation learning. Finally, we flexibly stack our blocks into a new powerful backbone, and adopt it for various vision tasks from image to video domain, from classification to dense prediction. Without any extra training data, our UniFormer achieves 86.3 top-1 accuracy on ImageNet-1 K classification task. With only ImageNet-1 K pre-training, it can simply achieve state-of-the-art performance in a broad range of downstream tasks. It obtains 82.9/84.8 top-1 accuracy on Kinetics-400/600, 60.9/71.2 top-1 accuracy on Something-Something V1/V2 video classification tasks, 53.8 box AP and 46.4 mask AP on COCO object detection task, 50.8 mIoU on ADE20 K semantic segmentation task, and 77.4 AP on COCO pose estimation task. Moreover, we build an efficient UniFormer with a concise hourglass design of token shrinking and recovering, which achieves 2-4[Formula: see text] higher throughput than the recent lightweight models.

SimpleClick: Interactive Image Segmentation with Simple Vision Transformers.

Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive image segmentation. To fill this gap, we propose SimpleClick, the first interactive segmentation method that leverages a plain backbone. Based on the plain backbone, we introduce a symmetric patch embedding layer that encodes clicks into the backbone with minor modifications to the backbone itself. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving 21.8% over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We provide a detailed computational analysis, highlighting the suitability of our method as a practical annotation tool.

What Makes for Hierarchical Vision Transformer?

Recent studies indicate that hierarchical Vision Transformer (ViT) with a macro architecture of interleaved non-overlapped window-based self-attention & shifted-window operation can achieve state-of-the-art performance in various visual recognition tasks, and challenges the ubiquitous convolutional neural networks (CNNs) using densely slid kernels. In most recently proposed hierarchical ViTs, self-attention is the de-facto standard for spatial information aggregation. In this paper, we question whether self-attention is the only choice for hierarchical ViT to attain strong performance, and study the effects of different kinds of cross-window communication methods. To this end, we replace self-attention layers with embarrassingly simple linear mapping layers, and the resulting proof-of-concept architecture termed TransLinear can achieve very strong performance in ImageNet-[Formula: see text] image recognition. Moreover, we find that TransLinear is able to leverage the ImageNet pre-trained weights and demonstrates competitive transfer learning properties on downstream dense prediction tasks such as object detection and instance segmentation. We also experiment with other alternatives to self-attention for content aggregation inside each non-overlapped window under different cross-window communication approaches. Our results reveal that the macro architecture, other than specific aggregation layers or cross-window communication mechanisms, is more responsible for hierarchical ViT's strong performance and is the real challenger to the ubiquitous CNN's dense sliding window paradigm.

SwinCross: Cross-modal Swin transformer for head-and-neck tumor segmentation in PET/CT images.

Radiotherapy (RT) combined with cetuximab is the standard treatment for patients with inoperable head and neck cancers. Segmentation of head and neck (H&N) tumors is a prerequisite for radiotherapy planning but a time-consuming process. In recent years, deep convolutional neural networks (DCNN) have become the de facto standard for automated image segmentation. However, due to the expensive computational cost associated with enlarging the field of view in DCNNs, their ability to model long-range dependency is still limited, and this can result in sub-optimal segmentation performance for objects with background context spanning over long distances. On the other hand, Transformer models have demonstrated excellent capabilities in capturing such long-range information in several semantic segmentation tasks performed on medical images. Despite the impressive representation capacity of vision transformer models, current vision transformer-based segmentation models still suffer from inconsistent and incorrect dense predictions when fed with multi-modal input data. We suspect that the power of their self-attention mechanism may be limited in extracting the complementary information that exists in multi-modal data. To this end, we propose a novel segmentation model, debuted, Cross-modal Swin Transformer (SwinCross), with cross-modal attention (CMA) module to incorporate cross-modal feature extraction at multiple resolutions. We propose a novel architecture for cross-modal 3D semantic segmentation with two main components: (1) a cross-modal 3D Swin Transformer for integrating information from multiple modalities (PET and CT), and (2) a cross-modal shifted window attention block for learning complementary information from the modalities. To evaluate the efficacy of our approach, we conducted experiments and ablation studies on the HECKTOR 2021 challenge dataset. We compared our method against nnU-Net (the backbone of the top-5 methods in HECKTOR 2021) and other state-of-the-art transformer-based models, including UNETR and Swin UNETR. The experiments employed a five-fold cross-validation setup using PET and CT images. Empirical evidence demonstrates that our proposed method consistently outperforms the comparative techniques. This success can be attributed to the CMA module's capacity to enhance inter-modality feature representations between PET and CT during head-and-neck tumor segmentation. Notably, SwinCross consistently surpasses Swin UNETR across all five folds, showcasing its proficiency in learning multi-modal feature representations at varying resolutions through the cross-modal attention modules. We introduced a cross-modal Swin Transformer for automating the delineation of head and neck tumors in PET and CT images. Our model incorporates a cross-modality attention module, enabling the exchange of features between modalities at multiple resolutions. The experimental results establish the superiority of our method in capturing improved inter-modality correlations between PET and CT for head-and-neck tumor segmentation. Furthermore, the proposed methodology holds applicability to other semantic segmentation tasks involving different imaging modalities like SPECT/CT or PET/MRI. Code:https://github.com/yli192/SwinCross_CrossModalSwinTransformer_for_Medical_Image_Segmentation.

A transformer-based neural ODE for dense prediction

A transformer-based neural ODE for dense prediction

An Information-Theoretic Method to Automatic Shortcut Avoidance and Domain Generalization for Dense Prediction Tasks.

Deep convolutional neural networks for dense prediction tasks are commonly optimized using synthetic data, as generating pixel-wise annotations for real-world data is laborious. However, the synthetically trained models do not generalize well to real-world environments. This poor "synthetic to real" (S2R) generalization we address through the lens of shortcut learning. We demonstrate that the learning of feature representations in deep convolutional networks is heavily influenced by synthetic data artifacts (shortcut attributes). To mitigate this issue, we propose an Information-Theoretic Shortcut Avoidance(ITSA) approach to automatically restrict shortcut-related information from being encoded into the feature representations. Specifically, our proposed method minimizes the sensitivity of latent features to input variations: to regularize the learning of robust and shortcut-invariant features in synthetically trained models. To avoid the prohibitive computational cost of direct input sensitivity optimization, we propose a practical yet feasible algorithm to achieve robustness. Our results show that the proposed method can effectively improve S2R generalization in multiple distinct dense prediction tasks, such as stereo matching, optical flow, and semantic segmentation. Importantly, the proposed method enhances the robustness of the synthetically trained networks and outperforms their fine-tuned counterparts (on real data) for challenging out-of-domain applications.