Robot manipulation pipelines, such as in bin-picking, often integrate perception with planning [1, 2, 3]. Some systems directly compute picks without computing the pose of objects by semantic segmentation or directly learning grasp affordances [4, 5, 6, 7]. While pose-agnostic techniques are promising, in many tasks it is important to first compute the 6D pose of observed objects to achieve purposeful manipulation and placement, such as in the context of packing [2, 8, 9].
Estimating 6D object poses has been approached in various ways , such as matching of locally-defined features , or of pre-defined templates of object models , or via voting in the local object frame using oriented point-pair features [13, 14]. Most methods were developed and evaluated for setups where each object appears once and for relatively sparsely placed objects on tabletops. Pose estimation for multiple instances of the same object type and where objects may be densely packed or in highly unstructured but dense piles has received less attention despite its significance in application domains, such as logistics. This is partly due to the increased difficulty of such setups.
There is prior work  that provides a public dataset  with a considerable number of instances for the same object type. Nevertheless, it measures the recall for estimating pose of any object instance in the scene. This may be sufficient in certain tasks but it is a weaker requirement than identifying the 6D pose of most, if not all, object instances. Achieving scene-level pose estimation allows a robot to internally simulate the world, reason about the order with which objects can be manipulated, as well as their physical interactions and the stability of their configuration. Scene-level reasoning can also infer missing information by considering occlusions and physical interactions between objects.
The current work aims to improve the robustness of pose estimation for real-world applications, where object types appear multiple times, and in challenging, dense configurations, such as those illustrated in Figure 1, while allowing for scene-level reasoning. It aims to do so by proposing effective machine learning operations that depend less on manual data labeling and less on handcrafted combinations of multiple accuracy criteria into a single objective function. This allows the true automation of robot manipulation pipelines and brings the hope of wider-scale, real-world deployment.
Problem Setup: The considered framework receives as input: a) an RGB image and a depth map of the scene; b) a set of mesh models , one for each object type present in the scene, and c) a set expressing an upper-bound on the number of instances for each object type . The output is object poses as a set of rigid-body transformations , where each captures the translation and rotation of the instance of object type in the camera’s reference frame, for each of the object types present in the scene.
Figure 1 summarizes the considered pipeline: a) CNNs are used to detect semantic object classes and visible boundaries of individual instances, b) then, a large set of candidate 6D pose hypotheses are generated for each object class, c) quality scores are computed for each hypothesis, and d) scene-level reasoning identifies consistent poses that maximize the sum of individual scores. In the context of this pipeline, the contribution of this work relative to state-of-the-art methods is two-fold:
A. Adversarial training with synthetic data for robust object class and boundary prediction: Machine learning approaches have become popular in pose estimation, both in end-to-end learning [16, 17] and as a pipeline component [18, 19]. They require, however, large amounts of labeled data. Recent approaches aim to solve single-instance pose estimation by training entirely in simulation [20, 19, 21]. The proposed method also utilizes labeled data generated exclusively in simulation to train a CNN for semantic segmentation. Nevertheless, CNNs are sensitive to the domain gap between synthetic and real data. The proposed training aims to mimic the physics of real-world test scenes and bridges the domain gap by using a generative adversarial network, as explained in Section 2. A key insight to improve robustness in the multi-instance case is that the network is simultaneously trained to predict object visibility boundaries. A thesis of this work is that boundaries learned on RGB images are more effective than boundaries detected on depth-maps  to guide and constrain the search for 6D poses, especially in tightly-packed setups.
B. Scene-level reasoning by automatically learning to evaluate pose candidate quality: This work finds the best physically-consistent set of poses among multiple candidates by formulating a constraint optimization problem and applying an ILP solver as shown in Section 3. The objective is to select pose hypotheses that maximize the sum of their individual scores, while respecting constraints, such as avoiding perceived collisions. Scene-level optimization has been previously approached as maximizing the weighted sum of various geometric features . The weights characterizing the objective function, however, were carefully handcrafted. This work shows it is possible to learn the distance of a given candidate pose from a ground-truth one by using a set of various objective functions as features. A gradient boosted tree  is trained to automatically integrate these objectives and regress the distance to the closest ground truth pose. The objectives indicate how well a candidate hypothesis explain the predicted object segments, the predicted boundaries, the observed depth and local surface normals in the input data. Prior related work has used tree search [24, 25, 26]
to reconstruct the scene by sequentially placing objects. These prior approaches, however, were restricted to a small number of objects or fewer degrees-of-freedom due to computational overhead. The proposedILP solution is quite fast in practice and scales to a large number of objects. For the images of Figure 1, the scene-level optimization is achieved in a few milliseconds.
2 Semantic and Boundary Predictions
Fully Convolutional Networks (FCNs) [27, 28] are popular semantic segmentation tools. They have also been used for object contour detection  and predicting multiple instances of an object type [30, 31]. These networks are increasingly being trained in simulation [32, 33, 34, 20] to alleviate the need for large amounts of labeled data. The domain gap between the data generated in simulation and real data can lead to noisy predictions and greatly affect pose estimation accuracy. Several recent methods have been developed to bridge this domain gap [35, 36]. The current work subscribes to these ideas and: a) exploits the constraints available in robotic setups to simulate scenes with realistic poses, while b) uses adversarial training with unlabeled real images to bridge the gap between the labels predicted in synthetic data with those predicted in real ones.
This work proposes the use of a CNN to predict per-pixel semantic classification and a classification of whether a pixel is a visible object boundary. The data for training the CNN are generated in simulation with a physics engine and a renderer. The simulation samples a bin pose and a camera pose given the robot’s workspace. Each scene is created by randomly sampling, within a pre-specified domain, the number and 6D poses of objects, the color of the bin, and the placement and intensity of the illumination sources. Finally, the scene is rendered to obtain a color image, a depth map, per-pixel class labels and visible instance boundary labels. The simulation generates a wide range of training data for domain randomization and robustness to domain gap issues. Nevertheless, domain gap still exists between synthetic data and data acquired through real sensors as it is hard to capture the domain of object material’s interactions with various illumination sources in the environment.
The generative adversarial network (GAN), shown in Figure 2, performs the semantic and boundary detection tasks, while also adapting the output predictions on unlabeled real images to resemble the predictions on synthetic images. It consists of a shared VGG16
encoder that stacks five blocks of convolution, batch normalization and max pooling layers. The network branches out into two decoders for semantic and boundary classifications. These are fully convolutional decoders with unpooling indices passed from corresponding max pooling blocks in the encoder section. The outputs of both decoders are passed to a corresponding fully convolutional discriminator network.
The network is trained by taking as input a synthetic image and its ground-truth label . It also receives as input unlabeled real image . The output of the label prediction network on image (or corresponding output on ) is then passed on to the label discriminator
whose task is to classify correctly if the prediction is on real or on synthetic data. Along with the objective of correctly labeling the synthetic image, the label prediction network should also confuse into classifying as an output of a sample coming from the synthetic domain. The objective of the semantic labeling network is defined as:
where is the per-pixel K-channel (for K object classes) output from the labeling network, are pixel coordinates, and is a weight factor. is the predicted score of in pixel of being a synthetic image. The domain classifier’s objective is specified as:
A similar GAN objective is used to simultaneously train the boundary predictor and discriminator.
3 Scene-level Pose Selection
This section formulates the scene-level objective and presents how to address it in a computationally efficient manner, despite its computational hardness. Given the semantic and boundary predictions, a set of 6D pose hypotheses is generated, , where denotes the number of hypotheses for each object class (). Pose candidates can be generated via learning , RANSAC  hough voting , or by incorporating distinctive geometric features .
) from the observed point cloud such that the points in each set belong with high probability to a single object instance. The sampled point sets are then matched to congruent sets on corresponding object model to generate candidate rigid transformations. A key feature of our hypotheses generation process, compared to[38, 19], is that the boundary predictions from the previous step are utilized to limit the selection of points within a single object instance. An adaptive sampling process is used to cover all instances of the same object by enforcing dispersion. The detailed formulation for the hypotheses generation process can be found in Appendix A.
The final objective is to select for an object category , a subset of poses , which maximizes the total sum of scores while avoiding collisions between objects when assigned to those poses. In other terms, the intersection between the volume occupied by any two object instances in the scene should be empty. The approach first identifies a set that contains all pairs of poses, which conflict with each other. This set is defined as where and is the volume occupied by a model when placed according to pose . is the maximum volume of tolerated collisions between objects in the scene. This positive error term is necessary because the best poses among the sampled ones may induce slight collisions between objects, which can often be corrected afterwards.
Pose selection is formulated as a constraint optimization problem, which can be solved by ILP
. A set of binary variablesis defined where . Each variable can be seen as an indicator variable on whether pose hypothesis is included in the set of selected posed . Then, the optimization problem is as follows:
The first constraint ensures that the number of poses selected for each category of object do not exceed the number of instances. The second constraint ensures that poses that are conflicting cannot be selected. This problem is equivalent to the Maximum-Weight Independent Set Problem (MWISP), with an additional constraint related to the number of selected poses. MWISP is NP-hard, and there are no -approximations for any fixed where is the number of variables . In practice, however, and for the problems considered here, an exact solution can be found very fast (in milliseconds) using modern ILP solvers for scenes containing a few hundreds of candidate poses. This is because the poses tend to cluster into cliques around specific instances, with a small number of constraints between poses in different cliques. Moreover, the objective function is monotone sub-modular as it is a linear function of a subset . Therefore, greedy optimization is guaranteed to find in linear time a solution that is at least fraction of the optimal. An approximate solution can also be found in linear time with an LP relaxation. After solving the ILP, the poses are constructed from by keeping those for which .
4 Pose Hypothesis Quality Evaluation
|: Visible boundary pixels when the object model is placed at pose and rendered|
|: Scene boundary pixels predicted by the CNN.|
|: Visible portion of the the object model when placed at pose and rendered.|
|: Surface and boundary matching distance thresholds.|
|: Depth distance between rendered image and observed depth image at pixel .|
|: Distance between rendered boundary and predicted boundary at pixel .|
|Model-to-Scene consistency features|
|: fraction of pixels in model boundary that match the scene boundary.|
|: fraction of pixels in scene boundary within the visible model region that match the model boundary.|
|: fraction of pixels in visible model region that is sufficiently aligned in terms of depth to the observed data.|
|Scene alignment features|
|: surface alignment score weighted by the corresponding label probability. Similarity score given by and it considers depth distance and surface normal similarity.|
|: boundary alignment score based on distance between point on model boundary and it’s nearest point in the predicted boundary set.|
Scene-level optimization, presented in Section 3, requires that a single quality score is assigned to each pose in the hypotheses set. The proposed score function considers five indicators of a good alignment for each pose candidate , shown in Table 1. A straightforward solution is to define as a weighted sum of its components. Nevertheless, the resulting poses would heavily depend on the choice of the weights. Choosing the right weights manually for every new object type is not trivial. Instead, a key aspect of this work is to learn the objective function using as features, i.e., . The function is learned by minimizing the following loss:
where is the distance between a given pose and a ground-truth pose . The distance is computed using the ADI metric, which is frequently used in the literature for evaluating pose estimations . This metric is explained in the experimental section.
The training data set is collected by simulating different scenes in a physics engine in the same way as described in Section. 2. For each scene, the CNN predicts semantic and object boundaries and a large number of pose hypotheses are sampled. For each hypothesis T, their alignment features and corresponding scores are computed based on the closest ground-truth pose. The regression learning problem is to find a function that maps features of a pose to its actual distance from the corresponding ground-truth pose .
This work adopts the Gradient Boosted Regression Trees (GBRT) to solve the optimization problem in Equation 1. GBRTs are well-suited for handling variables that have heterogeneous features . GBRTs are also a flexible non-parametric approach that can adapt to non-smooth changes in the regressed function using limited data, which often occurs when dealing with objects that have different shapes and sizes. An implementation of GBRT is available in the Scikit-learn library .
This section describes the experimental study performed on 2 datasets of scenes with multiple instances of objects. In this study, the recall for pose estimation is measured based on the error given by ADI , which measures distances between poses and given an object mesh model :
where corresponds to point after applying transformation on . Given a ground-truth pose , a true positive is a returned pose that has , where is a fraction set to , while is the diameter of the object model calculated as the maximum distance between any two points on the model.
The two datasets include a public dataset called bin-picking dataset  and a new one developed as part of this submission. The bin-picking dataset contains two object types and 3 scenarios (Figure 4). These scenarios present clutter of objects with high occlusion rate. For this dataset, segmenting objects in color images is challenging but depth cues can be used to find the object boundaries. This leads to depth-based approaches achieving high estimation success on this dataset without color information. The new dataset, henceforth called the densely-packed dataset, comprises of unique scenes with two object categories. Each scene contains to different instances of these objects, which were manually labeled with 6D pose annotations. There are two types of scenes in this dataset, as illustrated below. In one type, denoted as scenario 1, the instances are tightly packed next to each other. This case is particularly challenging because the surfaces of multiple instances are aligned, which makes it difficult to use depth information for segmentation. The dataset and the code is shared alongside the paper.
|Hinterstoisser at. al ||37.0||65.6||49.3|
|Buch at. al ||11.2||31.7||19.9|
|MRCNN-StoCS [43, 19]||42.8||68.3||53.7|
|PoseCNN + ICP ||56.8||80.6||67.0|
|Buch at. al ||63.8||44.9|
|Tejani et. al. ||31.4||24.8|
5.1 Evaluation against recent pose estimation techniques
Several state-of-the-art pose estimation techniques are evaluated on the above datasets (Table 4 and Table 4). A popular template-matching work  matches templates extracted from RGB and depth rendering of CAD models. It fails on several occasions as the templates are not robust to occlusion and varying lighting conditions. Approaches based on hough voting with point-pair features [13, 42] achieve high success on the bin-picking dataset but fail to do so on the densely-packed dataset. These approaches detect multiple object instances by considering the peaks in hough voting space, several of which are false positives by virtue of aligned surfaces in the packed boxes scenario. LCHF  is tailored to handle multiple instances of the same object category. Even after carefully tuning the weights of the optimization function and relaxing the criteria for the number of pose candidates to select, the recall from this approach is rather low. This can be majorly attributed to differences in pre-defined template descriptors between the scene and the object model used for matching local patches. LCHF  also includes an active vision component not considered in this evaluation.
Next, recent deep-learning based techniques for pose estimation are evaluated.StoCS  matches point cloud object models to stochastic output from a CNN trained for semantic segmentation with synthetic data. One way to apply StoCS for multi-instance estimation is to integrate it with an instance segmentation technique, such as Mask R-CNN . Mask R-CNN was trained with synthetic data and used to extract top instances of each object type according to the detection probability. Pose estimation was performed using StoCS for each of these individual segments. There are two major limitations of this combination. The first is the use of deterministic instance boundaries leading to segmentation noise that cannot be recovered during pose estimation. The second is that visibility is not considered when matching the model to the detected segment, which can lead to several incorrect poses achieving high alignment scores.
PoseCNN  is an end-to-end learning approach. It includes a network branch for semantic segmentation. Pixels belonging to an object class then vote for the object centroid’s location. Based on the peaks in voting, the center is localized and corresponding inliers are used to find a region of interest (RoI). Features of the RoI regress in a separate branch of the network to output the object’s rotation. PoseCNN was originally developed for single instances but via non-maximal suppression over the output of hough voting, it can be adapted for multiple instances. To eliminate domain gap from the scope of testing, PoseCNN was tested on a synthetic dataset with the same scenarios as the ones in real testing. PoseCNN outputs an object pose that could be used as an initialization for a depth-based ICP-like process that utilizes perturbations and local search for refinement. Overall, object symmetry and tight-packing scenarios make the simultaneous training of the multiple network branches hard to converge and the final ICP process less effective in these scenarios.
DOPE  is another learning-based approach that recovers 6D pose via perspective-n-point (PnP
) from predictions of 3D bounding-box vertices projected on the image. It aims to bridge the simulation-to-reality gap by a combination of domain randomization and photo-realistic rendering. The open-sourced rendering engineNDDS  used in DOPE, however, does not provide access to photo-realistic rendering. Thus, the comparisons were made against the DOPE-DR
version. The neural network output is composed of abelief map used to find the projected vertices by a local peak search as well as an affinity map, which indicates the direction from projected vertices to their corresponding center for assignment. Nevertheless, when multiple instances of the same object are placed next to each other, some 2D vertices significantly overlap around the border of two neighboring instances. This makes the assignment of vertices to the correct center problematic and degrade the performance of (PnP), since it requires relatively precise 2D-3D point correspondences. DOPE uses only color information without depth, which is a disadvantage when compared to other methods in Table 4. DOPE was trained from scratch with synthetic data generated by following the same pipeline as presented in . Using the best tuned parameters for domain randomization and post-processing steps, the pose estimation recall was measured on a synthetic validation set and on the real test set.
5.2 Ablation study of the proposed technique
The CNN in Section 2 is trained with 20,000 scenes generated in simulation by randomly dropping objects in the bin. It is then fine-tuned with images rendered from simulated scenes of tightly-packed objects. The networks were trained using the Adam optimizer with an initial learning rate of . The weight for the GAN loss is set to . To handle the class imbalance in the boundary network, the ratio of boundary to non-boundary pixels was computed in every iteration of the training and used to weight the respective loss terms.
|No adaptation (rgb)||69.8||79.5||73.9|
|ILP Objective function||o1||o2||All|
|Manual (scene + model)||76.4||83.1||79.2|
Table 8 indicates that label-space adaptation is the most effective training strategy in this case and even more so when scenes that mimic the packing scenario were used. Training solely on synthetic data with no adaptation significantly reduces the performance. An unpaired image translation approach, Cycle-GAN , achieves comparable performance. But with no semantic constraints, it biases the transfer for dominant classes in the dataset (o2 in this case), and also cannot deal with background clutter. Figure 4 shows synthetic training images for different datasets and boundary predictions on real images of corresponding datasets. To evaluate the generalization capacity of the training process, the network was also trained for the Occluded-Linemod dataset  that contains 8 object classes and unseen background clutter. Even then, the training was able to predict boundaries of only concerned objects.
Given the predictions from CNN, the hypotheses generation process described in Appendix A finds for each object type, a set of pose candidates, which should be large enough to include the true poses and small enough to reduce the computation time. The number of candidates is affected by the number of sampled bases and pose clustering. Recall rates are shown in Figure 8 separately for all generated candidates and the ones selected by ILP. It also shows the effect of dispersion parameter , which is by default set to . Table 8 evaluates the contribution of the boundary reasoning in the hypothesis generation and shows that it has a significant impact on the overall recall.
Table 8 compares learned objective function for ILP vs manually-defined ones. One way to combine features defined in Section 3 is to consider only scene alignment scores () as in many point registration algorithms or alternatively use both scene alignment and model consistency scores (). An upper bound of performance with the given hypotheses set is established as the optimal recall when the true ADI distance from ground-truth is used in the optimization.
The overall computation time for the current sequential implementation of the approach ranges from 10s to 15s for estimating all (15 to 19) instances in a scene. The CNN predictions, the pose hypotheses generation and the scene-level optimization along with collision checking run individually in less than a second. Broad phase collision checking is performed to speed up the process. The majority of computation time is spent on depth-buffering and local refinement for each pose candidate (130 per object category). These operations can be significantly sped up with parallel processing such as CUDA-OpenGL interoperability, which has been shown  to render 1000 images in 0.1s. Given the data parallelism, multiple cores can also be easily utilized to speed up the algorithm.
This work focuses on hard instances of scene-level, multi-instance pose estimation, which includes highly cluttered and densely packed scenarios. The results show that the type of poses used in simulation for training the semantic and the boundary networks is important. While a simulation-to-reality domain gap exists, it can be bridged by using appropriate information, such as boundary prediction, which translates well from simulated to real images, and adversarial training strategies. Furthermore, the ILP formulation of the scene-level reasoning is able to find combinations of hypotheses that are both consistent as well as of high-quality given the learned global score function. The consistency of pose hypotheses in the current work is defined by collision constraints. It is interesting to consider a learning process for identifying compatible sets of poses that express physical constraints and which could be used in the context of the proposed ILP formulation.
This work was supported by the NSF, grant numbers IIS-1734492 and IIS-1723869.
- Correll et al.  N. Correll, K. Bekris, D. Berenson, O. Brock, A. Causo, K. Hauser, K. Osada, A. Rodriguez, J. Romano, and P. Wurman. Analysis and Observations From the First Amazon Picking Challenge. T-ASE, 2016.
- Schwarz et al.  M. Schwarz, C. Lenz, G. García, S. Koo, A. Periyasamy, M. Schreiber, and S. Behnke. Fast object learning and dual-arm coordination for cluttered stowing, picking, and packing. In ICRA, 2018.
-  A. Zeng, K. Yu, S. Song, D. Suo, E. Walker, A. Rodriguez, and J. Xiao. Multiview self-supervised deep learning for 6d pose estimation in the amazon picking challenge. In ICRA’17.
- Gualtieri et al.  M. Gualtieri, A. Ten Pas, K. Saenko, and R. Platt. Grasp Pose Detection in Point Clouds. IJRR, 2017.
- Mahler et al.  J. Mahler, J. Liang, S. Niyaz, M. Laskey, R. Doan, X. Liu, J. Ojea, and K. Goldberg. Dex-net 2.0. In R:SS, 2017.
- Morrison et al.  D. Morrison et al. Cartman: The low-cost cartesian manipulator that won the amazon robotics challenge. In ICRA, 2018.
- Zeng et al.  A. Zeng et al. Robotic pick-and-place of novel objects in clutter with multi-affordance grasping and cross-domain image matching. In ICRA, 2018.
- Shome et al.  R. Shome, W. N. Tang, C. Song, C. Mitash, C. Kourtev, J. Yu, A. Boularias, and K. Bekris. Towards robust product packing with a minimalistic end-effector. In ICRA, 2019.
- Fan and Hauser  W. Fan and K. Hauser. Robot Packing with Known Items and Nondeterministic Arrival Order. In R:SS, 2019.
- Hodan et al.  T. Hodan, F. Michel, E. Brachmann, W. Kehl, A. GlentBuch, D. Kraft, B. Drost, J. Vidal, S. Ihrke, X. Zabulis, et al. Bop: benchmark for 6d object pose estimation. In ECCV, 2018.
- Aldoma et al.  A. Aldoma et al. Tutorial: Point cloud library: Three-dimensional object recognition and 6 dof pose estimation. IEEE Robotics & Automation Magazine, 2012.
- Hinterstoisser et al.  S. Hinterstoisser, V. Lepetit, S. Ilic, S. Holzer, G. Bradski, K. Konolige, and N. Navab. Model based training, detection and pose estimation of texture-less 3D objects in heavily cluttered scenes. In ACCV, 2012.
- Drost et al.  B. Drost, M. Ulrich, N. Navab, and S. Ilic. Model Globally, Match Locally: Efficient and Robust 3D Object Recognition. In CVPR, 2010.
- Vidal et al.  J. Vidal, C.-Y. Lin, X. Lladó, and R. Martí. A Method for 6D Pose Estimation of Free-Form Rigid Objects Using Point Pair Features on Range Data. Sensors, 2018.
- Doumanoglou et al.  A. Doumanoglou, R. Kouskouridas, S. Malassiotis, and T.-K. Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In CVPR, 2016.
Xiang et al. 
Y. Xiang, T. Schmidt, V. Narayanan, and D. Fox.
PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes.In R:SS, 2018.
- Kehl et al.  W. Kehl, F. Manhardt, F. Tombari, S. Ilic, and N. Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017.
- Brachmann et al.  E. Brachmann, A. Krull, F. Michel, S. Gumhold, J. Shotton, and C. Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014.
- Mitash et al.  C. Mitash, A. Boularias, and K. Bekris. Robust 6d object pose estimation with stochastic congruent sets. In BMVC, 2018.
- Tremblay et al.  J. Tremblay, T. To, B. Sundaralingam, Y. Xiang, D. Fox, and S. Birchfield. Deep Object Pose Estimation for Semantic Robotic Grasping of Household Objects. In CoRL, 2018.
- Sundermeyer et al.  M. Sundermeyer, Z.-C. Marton, M. Durner, M. Brucker, and R. Triebel. Implicit 3D Orientation Learning for 6D Object Detection from RGB Images. In ECCV, 2018.
- Aldoma et al.  A. Aldoma, F. Tombari, L. Di Stefano, and M. Vincze. A global hypotheses verification method for 3d object recognition. In ECCV, 2012.
-  J. Elith, J. R. Leathwick, and T. Hastie. A working guide to boosted regression trees. JAE’08.
- Narayanan and Likhachev  V. Narayanan and M. Likhachev. Discriminatively-guided Deliberative Perceptinon for Pose Estimation of Multiple 3D Object Instances. In R:SS, 2016.
- Sui et al.  Z. Sui, L. Xiang, O. C. Jenkins, and K. Desingh. Goal-directed robot manipulation through axiomatic scene estimation. In IJRR, 2017.
- Mitash et al.  C. Mitash, A. Boularias, and K. E. Bekris. Improving 6d pose estimation of objects in clutter via physics-aware monte carlo tree search. In ICRA, 2018.
- Long et al.  J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015.
- Noh et al.  H. Noh, S. Hong, and B. Han. Learning deconvolution network for semantic segmentation. In ICCV, 2015.
- Yang et al.  J. Yang, B. Price, S. Cohen, H. Lee, and M.-H. Yang. Object contour detection with a fully convolutional encoder-decoder network. In CVPR, 2016.
- Li et al.  Y. Li, H. Qi, J. Dai, X. Ji, and Y. Wei. Fully convolutional instance-aware semantic segmentation. In CVPR, 2017.
- Kirillov et al.  A. Kirillov, E. Levinkov, B. Andres, B. Savchynskyy, and C. Rother. Instancecut: from edges to instances with multicut. In CVPR, 2017.
Mitash et al. 
C. Mitash, K. Bekris, and A. Boularias.
A self-supervised learning system for object detection using physics simulation and multi-view pose estimation.In IROS, 2017.
- Tobin et al.  J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017.
- Hinterstoisser et al.  S. Hinterstoisser, V. Lepetit, P. Wohlhart, and K. Konolige. On pre-trained image features and synthetic images for deep learning. In ECCV, 2018.
Zhu et al. 
J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros.
Unpaired image-to-image translation using cycle-consistent adversarial networkss.In ICCV, 2017.
- Tsai et al.  Y.-H. Tsai, W.-C. Hung, S. Schulter, K. Sohn, M.-H. Yang, and M. Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018.
- Wang et al.  C. Wang, D. Xu, Y. Zhu, R. Martín-Martín, C. Lu, L. Fei-Fei, and S. Savarese. Densefusion: 6d object pose estimation by iterative dense fusion. In CVPR, 2019.
- Mellado et al.  N. Mellado, D. Aiger, and N. Mitra. Super4PCS: Fast Global Pointcloud Registration via Smart Indexing. In Computer Graphics Forum, 2014.
- Hastad  J. Hastad. Clique is hard to approximate within . In Acta Mathematica, 1996.
- Krause and Golovin  A. Krause and D. Golovin. Submodular Function Maximization. Cambridge Press, 2014.
- Pedregosa et al.  F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al. Scikit-learn: Machine learning in python. JMLR, 2011.
- Buch et al.  A. G. Buch, L. Kiforenko, and D. Kraft. Rotational subgroup voting and pose clustering for robust 3d object recognition. In ICCV, 2017.
- He et al.  K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask r-cnn. In ICCV, 2017.
- Tejani et al.  A. Tejani, D. Tang, R. Kouskouridas, and T. K. Kim. Latent-class Hough Forests for 3D Object Detection and Pose Estimation. In ECCV, 2014.
- To et al.  T. To, J. Tremblay, D. McKay, Y. Yamaguchi, K. Leung, A. Balanon, J. Cheng, W. Hodge, and S. Birchfield. NDDS: NVIDIA deep learning dataset synthesizer, 2018.