We propose an end-to-end Multitask Learning Transformer framework, named MulT, to simultaneously learn multiple high-level vision tasks, including depth estimation, semantic segmentation, reshading, surface normal estimation, 2D keypoint detection, and edge detection. Based on the Swin transformer model, our framework encodes the input image into a shared representation and makes predictions for each vision task using task-specific transformer-based decoder heads. At the heart of our approach is a shared attention mechanism modeling the dependencies across the tasks. We evaluate our model on several multitask benchmarks, showing that our MulT framework outperforms both the state-of-the art multitask convolutional neural network models and all the respective single task transformer models. Our experiments further highlight the benefits of sharing attention across all the tasks, and demonstrate that our MulT model is robust and generalizes well to new domains. Our project website is at https://ivrl.github.io/MulT/.
Self-supervised learning (SSL) methods aim to learn view-invariant representations by maximizing the similarity between the features extracted from different crops of the same image regardless of cropping size and content. In essence, this strategy ignores the fact that two crops may truly contain different image information, e.g., background and small objects, and thus tends to restrain the diversity of the learned representations. In this work, we address this issue by introducing a new self-supervised learning strategy, LoGo, that explicitly reasons about Local and Global crops. To achieve view invariance, LoGo encourages similarity between global crops from the same image, as well as between a global and a local crop. However, to correctly encode the fact that the content of smaller crops may differ entirely, LoGo promotes two local crops to have dissimilar representations, while being close to global crops. Our LoGo strategy can easily be applied to existing SSL methods. Our extensive experiments on a variety of datasets and using different self-supervised learning frameworks validate its superiority over existing approaches. Noticeably, we achieve better results than supervised models on transfer learning when using only 1/10 of the data.
Finding accurate correspondences among different views is the Achilles' heel of unsupervised Multi-View Stereo (MVS). Existing methods are built upon the assumption that corresponding pixels share similar photometric features. However, multi-view images in real scenarios observe non-Lambertian surfaces and experience occlusions. In this work, we propose a novel approach with neural rendering (RC-MVSNet) to solve such ambiguity issues of correspondences among views. Specifically, we impose a depth rendering consistency loss to constrain the geometry features close to the object surface to alleviate occlusions. Concurrently, we introduce a reference view synthesis loss to generate consistent supervision, even for non-Lambertian surfaces. Extensive experiments on DTU and Tanks\&Temples benchmarks demonstrate that our RC-MVSNet approach achieves state-of-the-art performance over unsupervised MVS frameworks and competitive performance to many supervised methods.The trained models and code will be released at https://github.com/Boese0601/RC-MVSNet.
We propose ways to obtain robust models against adversarial attacks from randomly-initialized binary networks. Unlike adversarial training, which learns the model parameters, we in contrast learn the structure of the robust model by pruning a randomly-initialized binary network. Our method confirms the strong lottery ticket hypothesis in the presence of adversarial attacks. Compared to the results obtained in a non-adversarial setting, we in addition improve the performance and compression of the model by 1) using an adaptive pruning strategy for different layers, and 2) using a different initialization scheme such that all model parameters are initialized either to +1 or -1. Our extensive experiments demonstrate that our approach performs not only better than the state-of-the art for robust binary networks; it also achieves comparable or even better performance than full-precision network training methods.
Deep image denoisers achieve state-of-the-art results but with a hidden cost. As witnessed in recent literature, these deep networks are capable of overfitting their training distributions, causing inaccurate hallucinations to be added to the output and generalizing poorly to varying data. For better control and interpretability over a deep denoiser, we propose a novel framework exploiting a denoising network. We call it controllable confidence-based image denoising (CCID). In this framework, we exploit the outputs of a deep denoising network alongside an image convolved with a reliable filter. Such a filter can be a simple convolution kernel which does not risk adding hallucinated information. We propose to fuse the two components with a frequency-domain approach that takes into account the reliability of the deep network outputs. With our framework, the user can control the fusion of the two components in the frequency domain. We also provide a user-friendly map estimating spatially the confidence in the output that potentially contains network hallucination. Results show that our CCID not only provides more interpretability and control, but can even outperform both the quantitative performance of the deep denoiser and that of the reliable filter, especially when the test data diverge from the training data.
Adversarial training is a popular method to robustify models against adversarial attacks. However, it exhibits much more severe overfitting than training on clean inputs. In this work, we investigate this phenomenon from the perspective of training instances, i.e., training input-target pairs. Based on a quantitative metric measuring instances' difficulty, we analyze the model's behavior on training instances of different difficulty levels. This lets us show that the decay in generalization performance of adversarial training is a result of the model's attempt to fit hard adversarial instances. We theoretically verify our observations for both linear and general nonlinear models, proving that models trained on hard instances have worse generalization performance than ones trained on easy instances. Furthermore, we prove that the difference in the generalization gap between models trained by instances of different difficulty levels increases with the size of the adversarial budget. Finally, we conduct case studies on methods mitigating adversarial overfitting in several scenarios. Our analysis shows that methods successfully mitigating adversarial overfitting all avoid fitting hard adversarial instances, while ones fitting hard adversarial instances do not achieve true robustness.
Generative Adversarial Network (GAN) based localized image editing can suffer ambiguity between semantic attributes. We thus present a novel objective function to evaluate the locality of an image edit. By introducing the supervision from a pre-trained segmentation network and optimizing the objective function, our framework, called Locally Effective Latent Space Direction (LELSD), is applicable to any dataset and GAN architecture. Our method is also computationally fast and exhibits a high extent of disentanglement, which allows users to interactively perform a sequence of edits on an image. Our experiments on both GAN-generated and real images qualitatively demonstrate the high quality and advantages of our method.
Estimating the depth of comics images is challenging as such images a) are monocular; b) lack ground-truth depth annotations; c) differ across different artistic styles; d) are sparse and noisy. We thus, use an off-the-shelf unsupervised image to image translation method to translate the comics images to natural ones and then use an attention-guided monocular depth estimator to predict their depth. This lets us leverage the depth annotations of existing natural images to train the depth estimator. Furthermore, our model learns to distinguish between text and images in the comics panels to reduce text-based artefacts in the depth estimates. Our method consistently outperforms the existing state-ofthe-art approaches across all metrics on both the DCM and eBDtheque images. Finally, we introduce a dataset to evaluate depth prediction on comics.
Image classification has significantly improved using deep learning. This is mainly due to convolutional neural networks (CNNs) that are capable of learning rich feature extractors from large datasets. However, most deep learning classification methods are trained on clean images and are not robust when handling noisy ones, even if a restoration preprocessing step is applied. While novel methods address this problem, they rely on modified feature extractors and thus necessitate retraining. We instead propose a method that can be applied on a pretrained classifier. Our method exploits a fidelity map estimate that is fused into the internal representations of the feature extractor, thereby guiding the attention of the network and making it more robust to noisy data. We improve the noisy-image classification (NIC) results by significantly large margins, especially at high noise levels, and come close to the fully retrained approaches. Furthermore, as proof of concept, we show that when using our oracle fidelity map we even outperform the fully retrained methods, whether trained on noisy or restored images.