Multi-task learning has recently emerged as a promising solution for a comprehensive understanding of complex scenes. In addition to being memory-efficient, multi-task models, when appropriately designed, can facilitate the exchange of complementary signals across tasks. In this work, we jointly address 2D semantic segmentation and three geometry-related tasks: dense depth estimation, surface normal estimation, and edge estimation, demonstrating their benefits on both indoor and outdoor datasets. We propose a novel multi-task learning architecture that leverages pairwise cross-task exchange through correlation-guided attention and self-attention to enhance the overall representation learning for all tasks. We conduct extensive experiments across three multi-task setups, showing the advantages of our approach compared to competitive baselines in both synthetic and real-world benchmarks. Additionally, we extend our method to the novel multi-task unsupervised domain adaptation setting. Our code is available at https://github.com/cv-rits/DenseMTL
Comprehensive 3D scene understanding, both geometrically and semantically, is important for real-world applications such as robot perception. Most of the existing work has focused on developing data-driven discriminative models for scene understanding. This paper provides a new approach to scene understanding, from a synthesis model perspective, by leveraging the recent progress on implicit 3D representation and neural rendering. Building upon the great success of Neural Radiance Fields (NeRFs), we introduce Scene-Property Synthesis with NeRF (SS-NeRF) that is able to not only render photo-realistic RGB images from novel viewpoints, but also render various accurate scene properties (e.g., appearance, geometry, and semantics). By doing so, we facilitate addressing a variety of scene understanding tasks under a unified framework, including semantic segmentation, surface normal estimation, reshading, keypoint detection, and edge detection. Our SS-NeRF framework can be a powerful tool for bridging generative learning and discriminative learning, and thus be beneficial to the investigation of a wide range of interesting problems, such as studying task relationships within a synthesis paradigm, transferring knowledge to novel tasks, facilitating downstream discriminative tasks as ways of data augmentation, and serving as auto-labeller for data creation.
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/.
Computer vision models excel at making predictions when the test distribution closely resembles the training distribution. Such models have yet to match the ability of biological vision to learn from multiple sources and generalize to new data sources and tasks. To facilitate the development and evaluation of more general vision systems, we introduce the General Robust Image Task (GRIT) benchmark. GRIT evaluates the performance, robustness, and calibration of a vision system across a variety of image prediction tasks, concepts, and data sources. The seven tasks in GRIT are selected to cover a range of visual skills: object categorization, object localization, referring expression grounding, visual question answering, segmentation, human keypoint detection, and surface normal estimation. GRIT is carefully designed to enable the evaluation of robustness under image perturbations, image source distribution shift, and concept distribution shift. By providing a unified platform for thorough assessment of skills and concepts learned by a vision model, we hope GRIT catalyzes the development of performant and robust general purpose vision systems.
Multi-task dense scene understanding is a thriving research domain that requires simultaneous perception and reasoning on a series of correlated tasks with pixel-wise prediction. Most existing works encounter a severe limitation of modeling in the locality due to heavy utilization of convolution operations, while learning interactions and inference in a global spatial-position and multi-task context is critical for this problem. In this paper, we propose a novel end-to-end Inverted Pyramid multi-task Transformer (InvPT) to perform simultaneous modeling of spatial positions and multiple tasks in a unified framework. To the best of our knowledge, this is the first work that explores designing a transformer structure for multi-task dense prediction for scene understanding. Besides, it is widely demonstrated that a higher spatial resolution is remarkably beneficial for dense predictions, while it is very challenging for existing transformers to go deeper with higher resolutions due to huge complexity to large spatial size. InvPT presents an efficient UP-Transformer block to learn multi-task feature interaction at gradually increased resolutions, which also incorporates effective self-attention message passing and multi-scale feature aggregation to produce task-specific prediction at a high resolution. Our method achieves superior multi-task performance on NYUD-v2 and PASCAL-Context datasets respectively, and significantly outperforms previous state-of-the-arts. The code is available at https://github.com/prismformore/InvPT
This work proposes a novel pose estimation model for object categories that can be effectively transferred to previously unseen environments. The deep convolutional network models (CNN) for pose estimation are typically trained and evaluated on datasets specifically curated for object detection, pose estimation, or 3D reconstruction, which requires large amounts of training data. In this work, we propose a model for pose estimation that can be trained with small amount of data and is built on the top of generic mid-level representations \cite{taskonomy2018} (e.g. surface normal estimation and re-shading). These representations are trained on a large dataset without requiring pose and object annotations. Later on, the predictions are refined with a small CNN neural network that exploits object masks and silhouette retrieval. The presented approach achieves superior performance on the Pix3D dataset \cite{pix3d} and shows nearly 35\% improvement over the existing models when only 25\% of the training data is available. We show that the approach is favorable when it comes to generalization and transfer to novel environments. Towards this end, we introduce a new pose estimation benchmark for commonly encountered furniture categories on challenging Active Vision Dataset \cite{Ammirato2017ADF} and evaluated the models trained on the Pix3D dataset.
This paper introduces a pipeline to parametrically sample and render multi-task vision datasets from comprehensive 3D scans from the real world. Changing the sampling parameters allows one to "steer" the generated datasets to emphasize specific information. In addition to enabling interesting lines of research, we show the tooling and generated data suffice to train robust vision models. Common architectures trained on a generated starter dataset reached state-of-the-art performance on multiple common vision tasks and benchmarks, despite having seen no benchmark or non-pipeline data. The depth estimation network outperforms MiDaS and the surface normal estimation network is the first to achieve human-level performance for in-the-wild surface normal estimation -- at least according to one metric on the OASIS benchmark. The Dockerized pipeline with CLI, the (mostly python) code, PyTorch dataloaders for the generated data, the generated starter dataset, download scripts and other utilities are available through our project website, https://omnidata.vision.
Surface normal estimation from a single image is an important task in 3D scene understanding. In this paper, we address two limitations shared by the existing methods: the inability to estimate the aleatoric uncertainty and lack of detail in the prediction. The proposed network estimates the per-pixel surface normal probability distribution. We introduce a new parameterization for the distribution, such that its negative log-likelihood is the angular loss with learned attenuation. The expected value of the angular error is then used as a measure of the aleatoric uncertainty. We also present a novel decoder framework where pixel-wise multi-layer perceptrons are trained on a subset of pixels sampled based on the estimated uncertainty. The proposed uncertainty-guided sampling prevents the bias in training towards large planar surfaces and improves the quality of prediction, especially near object boundaries and on small structures. Experimental results show that the proposed method outperforms the state-of-the-art in ScanNet and NYUv2, and that the estimated uncertainty correlates well with the prediction error. Code is available at https://github.com/baegwangbin/surface_normal_uncertainty.
Despite the fast progress in training specialized models for various tasks, learning a single general model that works well for many tasks is still challenging for computer vision. Here we introduce multi-task self-training (MuST), which harnesses the knowledge in independent specialized teacher models (e.g., ImageNet model on classification) to train a single general student model. Our approach has three steps. First, we train specialized teachers independently on labeled datasets. We then use the specialized teachers to label an unlabeled dataset to create a multi-task pseudo labeled dataset. Finally, the dataset, which now contains pseudo labels from teacher models trained on different datasets/tasks, is then used to train a student model with multi-task learning. We evaluate the feature representations of the student model on 6 vision tasks including image recognition (classification, detection, segmentation)and 3D geometry estimation (depth and surface normal estimation). MuST is scalable with unlabeled or partially labeled datasets and outperforms both specialized supervised models and self-supervised models when training on large scale datasets. Lastly, we show MuST can improve upon already strong checkpoints trained with billions of examples. The results suggest self-training is a promising direction to aggregate labeled and unlabeled training data for learning general feature representations.
This paper focuses on a new problem of estimating human pose and shape from single polarization images. Polarization camera is known to be able to capture the polarization of reflected lights that preserves rich geometric cues of an object surface. Inspired by the recent applications in surface normal reconstruction from polarization images, in this paper, we attempt to estimate human pose and shape from single polarization images by leveraging the polarization-induced geometric cues. A dedicated two-stage pipeline is proposed: given a single polarization image, stage one (Polar2Normal) focuses on the fine detailed human body surface normal estimation; stage two (Polar2Shape) then reconstructs clothed human shape from the polarization image and the estimated surface normal. To empirically validate our approach, a dedicated dataset (PHSPD) is constructed, consisting of over 500K frames with accurate pose and parametric shape annotations. Empirical evaluations on this real-world dataset as well as a synthetic dataset, SURREAL, demonstrate the effectiveness of our approach. It suggests polarization camera as a promising alternative to the more conventional RGB camera for human pose and shape estimation.
Freespace detection is a fundamental component of autonomous driving perception. Recently, deep convolutional neural networks (DCNNs) have achieved impressive performance for this task. In particular, SNE-RoadSeg, our previously proposed method based on a surface normal estimator (SNE) and a data-fusion DCNN (RoadSeg), has achieved impressive performance in freespace detection. However, SNE-RoadSeg is computationally intensive, and it is difficult to execute in real time. To address this problem, we introduce SNE-RoadSeg+, an upgraded version of SNE-RoadSeg. SNE-RoadSeg+ consists of 1) SNE+, a module for more accurate surface normal estimation, and 2) RoadSeg+, a data-fusion DCNN that can greatly minimize the trade-off between accuracy and efficiency with the use of deep supervision. Extensive experimental results have demonstrated the effectiveness of our SNE+ for surface normal estimation and the superior performance of our SNE-RoadSeg+ over all other freespace detection approaches. Specifically, our SNE-RoadSeg+ runs in real time, and meanwhile, achieves the state-of-the-art performance on the KITTI road benchmark. Our project page is at https://www.sne-roadseg.site/sne-roadseg-plus.
We review solutions to the problem of depth estimation, arguably the most important subtask in scene understanding. We focus on the single image depth estimation problem. Due to its properties, the single image depth estimation problem is currently best tackled with machine learning methods, most successfully with convolutional neural networks. We provide an overview of the field by examining key works. We examine non-deep learning approaches that mostly predate deep learning and utilize hand-crafted features and assumptions, and more recent works that mostly use deep learning techniques. The single image depth estimation problem is tackled first in a supervised fashion with absolute or relative depth information acquired from human or sensor-labeled data, or in an unsupervised way using unlabelled stereo images or video datasets. We also study multitask approaches that combine the depth estimation problem with related tasks such as semantic segmentation and surface normal estimation. Finally, we discuss investigations into the mechanisms, principles, and failure cases of contemporary solutions.
We present streaming self-training (SST) that aims to democratize the process of learning visual recognition models such that a non-expert user can define a new task depending on their needs via a few labeled examples and minimal domain knowledge. Key to SST are two crucial observations: (1) domain-agnostic unlabeled images enable us to learn better models with a few labeled examples without any additional knowledge or supervision; and (2) learning is a continuous process and can be done by constructing a schedule of learning updates that iterates between pre-training on novel segments of the streams of unlabeled data, and fine-tuning on the small and fixed labeled dataset. This allows SST to overcome the need for a large number of domain-specific labeled and unlabeled examples, exorbitant computational resources, and domain/task-specific knowledge. In this setting, classical semi-supervised approaches require a large amount of domain-specific labeled and unlabeled examples, immense resources to process data, and expert knowledge of a particular task. Due to these reasons, semi-supervised learning has been restricted to a few places that can house required computational and human resources. In this work, we overcome these challenges and demonstrate our findings for a wide range of visual recognition tasks including fine-grained image classification, surface normal estimation, and semantic segmentation. We also demonstrate our findings for diverse domains including medical, satellite, and agricultural imagery, where there does not exist a large amount of labeled or unlabeled data.
While convolutional neural networks have shown a tremendous impact on various computer vision tasks, they generally demonstrate limitations in explicitly modeling long-range dependencies due to the intrinsic locality of the convolution operation. Initially designed for natural language processing tasks, Transformers have emerged as alternative architectures with innate global self-attention mechanisms to capture long-range dependencies. In this paper, we propose TransDepth, an architecture that benefits from both convolutional neural networks and transformers. To avoid the network losing its ability to capture local-level details due to the adoption of transformers, we propose a novel decoder that employs attention mechanisms based on gates. Notably, this is the first paper that applies transformers to pixel-wise prediction problems involving continuous labels (i.e., monocular depth prediction and surface normal estimation). Extensive experiments demonstrate that the proposed TransDepth achieves state-of-the-art performance on three challenging datasets. Our code is available at: https://github.com/ygjwd12345/TransDepth.
In this paper, we propose a geometric neural network with edge-aware refinement (GeoNet++) to jointly predict both depth and surface normal maps from a single image. Building on top of two-stream CNNs, GeoNet++ captures the geometric relationships between depth and surface normals with the proposed depth-to-normal and normal-to-depth modules. In particular, the "depth-to-normal" module exploits the least square solution of estimating surface normals from depth to improve their quality, while the "normal-to-depth" module refines the depth map based on the constraints on surface normals through kernel regression. Boundary information is exploited via an edge-aware refinement module. GeoNet++ effectively predicts depth and surface normals with strong 3D consistency and sharp boundaries resulting in better reconstructed 3D scenes. Note that GeoNet++ is generic and can be used in other depth/normal prediction frameworks to improve the quality of 3D reconstruction and pixel-wise accuracy of depth and surface normals. Furthermore, we propose a new 3D geometric metric (3DGM) for evaluating depth prediction in 3D. In contrast to current metrics that focus on evaluating pixel-wise error/accuracy, 3DGM measures whether the predicted depth can reconstruct high-quality 3D surface normals. This is a more natural metric for many 3D application domains. Our experiments on NYUD-V2 and KITTI datasets verify that GeoNet++ produces fine boundary details, and the predicted depth can be used to reconstruct high-quality 3D surfaces. Code has been made publicly available.
We present HoliCity, a city-scale 3D dataset with rich structural information. Currently, this dataset has 6,300 real-world panoramas of resolution $13312 \times 6656$ that are accurately aligned with the CAD model of downtown London with an area of more than 20 km$^2$, in which the median reprojection error of the alignment of an average image is less than half a degree. This dataset aims to be an all-in-one data platform for research of learning abstracted high-level holistic 3D structures that can be derived from city CAD models, e.g., corners, lines, wireframes, planes, and cuboids, with the ultimate goal of supporting real-world applications including city-scale reconstruction, localization, mapping, and augmented reality. The accurate alignment of the 3D CAD models and panoramas also benefits low-level 3D vision tasks such as surface normal estimation, as the surface normal extracted from previous LiDAR-based datasets is often noisy. We conduct experiments to demonstrate the applications of HoliCity, such as predicting surface segmentation, normal maps, depth maps, and vanishing points, as well as test the generalizability of methods trained on HoliCity and other related datasets. HoliCity is available at https://holicity.io.
With a proliferation of generic domain-adaptation approaches, we report a simple yet effective technique for learning difficult per-pixel 2.5D and 3D regression representations of articulated people. We obtained strong sim-to-real domain generalization for the 2.5D DensePose estimation task and the 3D human surface normal estimation task. On the multi-person DensePose MSCOCO benchmark, our approach outperforms the state-of-the-art methods which are trained on real images that are densely labelled. This is an important result since obtaining human manifold's intrinsic uv coordinates on real images is time consuming and prone to labeling noise. Additionally, we present our model's 3D surface normal predictions on the MSCOCO dataset that lacks any real 3D surface normal labels. The key to our approach is to mitigate the "Inter-domain Covariate Shift" with a carefully selected training batch from a mixture of domain samples, a deep batch-normalized residual network, and a modified multi-task learning objective. Our approach is complementary to existing domain-adaptation techniques and can be applied to other dense per-pixel pose estimation problems.
In this paper, we present a spatial rectifier to estimate surface normals of tilted images. Tilted images are of particular interest as more visual data are captured by arbitrarily oriented sensors such as body-/robot-mounted cameras. Existing approaches exhibit bounded performance on predicting surface normals because they were trained using gravity-aligned images. Our two main hypotheses are: (1) visual scene layout is indicative of the gravity direction; and (2) not all surfaces are equally represented by a learned estimator due to the structured distribution of the training data, thus, there exists a transformation for each tilted image that is more responsive to the learned estimator than others. We design a spatial rectifier that is learned to transform the surface normal distribution of a tilted image to the rectified one that matches the gravity-aligned training data distribution. Along with the spatial rectifier, we propose a novel truncated angular loss that offers a stronger gradient at smaller angular errors and robustness to outliers. The resulting estimator outperforms the state-of-the-art methods including data augmentation baselines not only on ScanNet and NYUv2 but also on a new dataset called Tilt-RGBD that includes considerable roll and pitch camera motion.
We present a case-study to improve the task-specific representation by leveraging a million unlabelled images without any extra knowledge. We propose an exceedingly simple method of conditioning an existing representation on a diverse data distribution and observe that a model trained on diverse examples acts as a better initialization. We extensively study our findings for the task of surface normal estimation and semantic segmentation from a single image. We improve surface normal estimation on NYU-v2 depth dataset and semantic segmentation on PASCAL VOC by 4% over base model. We did not use any task-specific knowledge or auxiliary tasks, neither changed hyper-parameters nor made any modification in the underlying neural network architecture.
We develop new algorithms for simultaneous learning of multiple tasks (e.g., image classification, depth estimation), and for adapting to unseen task/domain distributions within those high-level tasks (e.g., different environments). First, we learn common representations underlying all tasks. We then propose an attention mechanism to dynamically specialize the network, at runtime, for each task. Our approach is based on weighting each feature map of the backbone network, based on its relevance to a particular task. To achieve this, we enable the attention module to learn task representations during training, which are used to obtain attention weights. Our method improves performance on new, previously unseen environments, and is 1.5x faster than standard existing meta learning methods using similar architectures. We highlight performance improvements for Multi-Task Meta Learning of 4 tasks (image classification, depth, vanishing point, and surface normal estimation), each over 10 to 25 test domains/environments, a result that could not be achieved with standard meta learning techniques like MAML.