Label assignment in object detection aims to assign targets, foreground or background, to sampled regions in an image. Unlike labeling for image classification, this problem is not well defined due to the object’s bounding box. In this paper, we investigate the problem from a perspective of distillation, hence we call Label Assignment Distillation (LAD). Our initial motivation is very simple, we use a teacher network to generate labels for the student. This can be achieved in two ways: either using the teacher’s prediction as the direct targets (soft label), or through the hard labels dynamically assigned by the teacher (LAD). Our experiments reveal that: (i) LAD is more effective than soft-label, but they are complementary. (ii) Using LAD, a smaller teacher can also improve a larger student significantly, while soft-label can’t. We then introduce Co-learning LAD, in which two networks simultaneously learn from scratch and the role of teacher and student are dynamically interchanged. Using PAA-ResNet50 as a teacher, our LAD techniques can improve detectors PAA-ResNet101 and PAA-ResNeXt101 to 46AP and 47.5AP on the COCO test-dev set. With a stronger teacher PAA-SwinB, we improve the students PAA-ResNet50 to 43.7AP by only 1× schedule training and standard setting, and PAA-ResNet101 to 47.9AP, significantly surpassing the current methods. Our source code is released at https://git.io/JrDZo.

This paper addresses the problem of common object detection, which aims to detect objects of similar categories from a set of images. Although it shares some similarities with the standard object detection and co-segmentation, common object detection, recently promoted by \cite{Jiang2019a}, has some unique advantages and challenges. First, it is designed to work on both closed-set and open-set conditions, a.k.a. known and unknown objects. Second, it must be able to match objects of the same category but not restricted to the same instance, texture, or posture. Third, it can distinguish multiple objects. In this work, we introduce the Single Stage Common Object Detection (SSCOD) to detect class-agnostic common objects from an image set. The proposed method is built upon the standard single-stage object detector. Furthermore, an embedded branch is introduced to generate the object's representation feature, and their similarity is measured by cosine distance. Experiments are conducted on PASCAL VOC 2007 and COCO 2014 datasets. While being simple and flexible, our proposed SSCOD built upon ATSSNet performs significantly better than the baseline of the standard object detection, while still be able to match objects of unknown categories. Our source code can be found at \href{https://github.com/cybercore-co-ltd/Single-Stage-Common-Object-Detection}{(URL)}

Taking an image of an object is at its core a lossy process. The rich information about the three-dimensional structure of the world is flattened to an image plane and decisions such as viewpoint and camera parameters are final and not easily revertible. As a consequence, possibilities of changing viewpoint are limited. Given a single image depicting an object, novel-view synthesis is the task of generating new images that render the object from a different viewpoint than the one given. The main difficulty is to synthesize the parts that are disoccluded; disocclusion occurs when parts of an object are hidden by the object itself under a specific viewpoint. In this work, we show how to improve novel-view synthesis by making use of the correlations observed in 3D models and applying them to new image instances. We propose a technique to use the structural information extracted from a 3D model that matches the image object in terms of viewpoint and shape. For the latter part, we propose an efficient 2D-to-3D alignment method that associates precisely the image appearance with the 3D model geometry with minimal user interaction. Our technique is able to simulate plausible viewpoint changes for a variety of object classes within seconds. Additionally, we show that our synthesized images can be used as additional training data that improves the performance of standard object detectors.

Abstract We propose a system to restrict the manipulation of shape and appearance in an image to a valid subspace which we learn from a collection of exemplar images. To this end, we automatically co‐align a collection of images and learn a subspace model of shape and appearance using principal components. As finding perfect image correspondences for general images is not feasible, we build an approximate partial alignment and improve bad alignments leveraging other, more successful alignments. Our system allows the user to change appearance and shape in real‐time and the result is “projected” onto the subspace of meaningful changes. The change in appearance and shape can either be locked or performed independently. Additional applications include suggestion of alternative shapes or appearance.

Color selection is required in many computer graphics applications, but can be tedious, as 1D or 2D user interfaces are employed to navigate in a 3D color space. Until now the problem was considered a question of designing general color spaces with meaningful (e.g., perceptual) parameters. In this work, we show how color selection usability improves by applying 1D or 2D color manifolds that predict the most likely change of color in a specific context. A typical use-case is manipulating the color of a banana; instead of presenting a 2D+1D RGB, CIE Lab, or HSV widget, our approach presents a simple 1D slider that captures the most likely change for this context. Technically, for each context, we learn a lower-dimensional manifold with varying density from labeled Internet examples. We demonstrate the increase in task performance of color selection in a user study.

Abstract This work addresses the challenge of intuitive appearance editing in scenes with complex geometric layout and complex, spatially‐varying indirect lighting. In contrast to previous work, that aimed to edit surface reflectance, our system allows a user to freely manipulate the surface light field. It then finds the best surface reflectance that “explains” the surface light field manipulation. Instead of classic L2fitting of reflectance to a combination of incoming and exitant illumination, our system infers a sparse L0change of shading parameters instead. Consequently, our system does not require “diffuse” or “glossiness” brushes or any such understanding of the underlying reflectance parametrization. Instead, it infers reflectance changes from scribbles made by a single simple color brush tool alone: Drawing a highlight will increase Phong specular; blurring a mirror reflection will decrease glossiness; etc. A sparse‐solver framework operating on a novel point‐based, pre‐convolved lighting representation in combination with screen‐space edit upsampling allows to perform editing interactively on a GPU.

Abstract This work proposes a technique to transfer the material style or mood from a guide source such as an image or video onto a target 3D scene. It formulates the problem as a combinatorial optimization of assigning discrete materials extracted from the guide source to discrete objects in the target 3D scene. The assignment is optimized to fulfill multiple goals: overall image mood based on several image statistics; spatial material organization and grouping as well as geometric similarity between objects that were assigned to similar materials. To be able to use common uncalibrated images and videos with unknown geometry and lighting as guides, a material estimation derives perceptually plausible reflectance, specularity, glossiness, and texture. Finally, results produced by our method are compared to manual material assignments in a perceptual study.

Abstract The incident indirect light over a range of image pixels is often coherent. Two common approaches to exploit this inter‐pixel coherence to improve rendering performance are Irradiance Caching and Radiance Caching. Both compute incident indirect light only for a small subset of pixels (the cache), and later interpolate between pixels. Irradiance Caching uses scalar values that can be interpolated efficiently, but cannot account for shading variations caused by normal and reflectance variation between cache items. Radiance Caching maintains directional information, e.g., to allow highlights between cache items, but at the cost of storing and evaluating a Spherical Harmonics (SH) function per pixel. The arithmetic and bandwidth cost for this evaluation is linear in the number of coefficients and can be substantial. In this paper, we propose a method to replace it by an efficient per‐cache item pre‐filtering based on MIP maps — such as previously done for environment maps — leading to a single constant‐time lookup per pixel. Additionally, per‐cache item geometry statistics stored in distance‐MIP maps are used to improve the quality of each pixel's lookup. Our approximate interactive global illumination approach is an order of magnitude faster than Radiance Caching with Phong BRDFs and can be combined with Monte Carlo‐raytracing, Point‐based Global Illumination or Instant Radiosity.

Abstract We propose a novel rendering method which supports interactive BRDF editing as well as relighting on a 3D scene. For interactive BRDF editing, we linearize an analytic BRDF model with basis BRDFs obtained from a principal component analysis. For each basis BRDF, the radiance transfer is precomputed and stored in vector form. In rendering time, illumination of a point is computed by multiplying the radiance transfer vectors of the basis BRDFs by the incoming radiance from gather samples and then linearly combining the results weighted by user‐controlled parameters. To improve the level of accuracy, a set of sub‐area samples associated with a gather sample refines the glossy reflection of the geometric details without increasing the precomputation time. We demonstrate this program with a number of examples to verify the real‐time performance of relighting and BRDF editing on 3D scenes with complex lighting and geometry.