Back

Unified Perceptual Parsing for Scene Understanding

Tete XiaoYingcheng LiuBolei ZhouYuning JiangJian Sun

article2018ECCV2,638 citations

Develops UPerNet, a multi-task framework that learns from heterogeneous annotations to simultaneously parse visual concepts across scenes, objects, parts, and textures for comprehensive image understanding.

Listen

The article addresses the gap between human visual perception, which integrates multiple levels of understanding from scenes and objects down to parts, materials, and textures, and current computer vision systems that typically handle these tasks separately. This separation limits the ability of machines to achieve comprehensive scene understanding, which is increasingly important for applications like autonomous systems and image analysis.

The article set out to define and solve a new task called Unified Perceptual Parsing, which requires a single model to recognize as many visual concepts as possible from an image, and to develop an efficient framework for learning from diverse, incomplete datasets.

The authors combined several existing datasets into a standardized collection called Broden+ containing over 57,000 images with heterogeneous annotations. They designed UPerNet, a Feature Pyramid Network architecture with task-specific heads attached at different semantic levels, and used a sampling strategy that updates only relevant layers for each data source during training.

The framework achieved competitive results on object segmentation with a mean IoU of 41.22 percent on the ADE20K benchmark while using roughly one-third the training time of prior methods. Joint training across tasks yielded strong performance on parts, materials, and scenes, with only minor degradation on some metrics, and enabled the model to infer pixel-level texture labels from image-level annotations. The trained network also extracted meaningful hierarchical relations, such as common objects in indoor versus outdoor scenes and typical materials for specific objects and parts.

These findings show that a single network can unify multiple perceptual levels without prohibitive computational cost, supporting more complete scene understanding that could improve downstream tasks like reasoning and knowledge discovery. The results differ from expectations that multi-task learning on heterogeneous data would cause large performance drops, highlighting the value of hierarchical feature attachment.

The article recommends further work on better fusion of synthetic and natural image sources to improve texture prediction. Additional data and analysis would strengthen claims before deploying the approach in production systems.

The main limitations include reliance on a fixed set of datasets that may not cover all real-world variability, slight performance trade-offs from joint training, and suboptimal texture results from the current fine-tuning approach. Confidence in the core segmentation and multi-task results is high based on the reported benchmarks, but caution is warranted for texture-related applications until refinements are tested.

ByteDanceMassachusetts Institute of TechnologyMegvii TechnologyPeking Universityimage annotationmachine visionparsing algorithmUnified Perceptual ParsingUPerNetBroden+
arXiv: 1807.10221CSAILVision/unifiedparsing
Cover for Unified Perceptual Parsing for Scene Understanding

Abstract

Humans recognize the visual world at multiple levels: we effortlessly categorize scenes and detect objects inside, while also identifying the textures and surfaces of the objects along with their different compositional parts. In this paper, we study a new task called Unified Perceptual Parsing, which requires the machine vision systems to recognize as many visual concepts as possible from a given image. A multi-task framework called UPerNet and a training strategy are developed to learn from heterogeneous image annotations. We benchmark our framework on Unified Perceptual Parsing and show that it is able to effectively segment a wide range of concepts from images. The trained networks are further applied to discover visual knowledge in natural scenes. Models are available at \url{this https URL}.

Topics

image annotationmachine visionparsing algorithmscene understandingvision systems

Keywords

Unified Perceptual ParsingUPerNetBroden+Feature Pyramid NetworkPyramid Pooling ModuleMulti‑Task LearningScene UnderstandingSemantic SegmentationHeterogeneous AnnotationVisual Knowledge Discovery

Table of Contents

  • 1 Introduction
  • 1.1 Related work
  • 2 Defining Unified Perceptual Parsing
  • 2.1 Datasets
  • 2.2 Metrics
  • 3 Designing Networks for Unified Perceptual Parsing
  • 3.1 Implementation details
  • 3.2 Design discussion
  • 4 Experiments
  • 4.1 Main results
  • 4.2 Discovering visual knowledge in natural scenes
  • 5 Conclusion
  • References

Knowls

  1. Knowl 1 — Unified Perceptual Parsing Network (UPerNet) Architecture

    model/method

    The Unified Perceptual Parsing Network (UPerNet) is a multi-task convolutional network designed to parse visual concepts across multiple semantic and perceptual levels (scene, object, object part, material, and texture) from an image.

    The framework consists of:

    1. Backbone and Feature Pyramid: A Residual Network (ResNet) backbone outputs multi-stage feature maps denoted {C2,C3,C4,C5}\{C_2, C_3, C_4, C_5\} at down-sampling strides {4,8,16,32}\{4, 8, 16, 32\}.
    2. Pyramid Pooling Module (PPM): A PPM is attached to the final backbone feature map C5C_5 to capture global context priors before top-down feature propagation.
    3. Feature Pyramid Network (FPN): The PPM output and lower backbone features are processed through top-down lateral connections, yielding feature pyramid maps {P2,P3,P4,P5}\{P_2, P_3, P_4, P_5\} at strides {4,8,16,32}\{4, 8, 16, 32\}, where P5P_5 is the feature map directly after the PPM.
    4. Hierarchical Task Heads:
      • Scene Head: Attached directly to P5P_5. It applies global average pooling followed by a linear classification layer to output image-level scene categories.
      • Object and Part Heads: Attached to a fused multi-scale feature representation. The pyramid maps {P3,P4,P5}\{P_3, P_4, P_5\} are up-sampled to the spatial resolution of P2P_2 via bilinear interpolation, concatenated with P2P_2, and projected through a 1×11 \times 1 convolution (with 512 channels, batch normalization, and ReLU) before separate convolutional classification heads for objects and parts.
      • Material Head: Attached directly to the highest-resolution pyramid map P2P_2 (stride 4) to preserve fine-grained local appearance and surface boundary features.
      • Texture Head: Appended as several convolutional layers on top of early backbone features (C2C_2). It is fine-tuned after training other heads, with backbone gradients blocked, to predict localized pixel-wise texture labels.

  2. Knowl 2 — Unified Perceptual Parsing Task Definition

    definition

    Unified Perceptual Parsing (UPP) is a computer vision task requiring a model to parse multiple visual concepts across perceptual hierarchies from a single input image simultaneously. The visual concepts span five distinct perceptual levels:

    1. Scene Classification: Predicting holistic, image-level category labels (e.g., living room, runway).
    2. Object Semantic Segmentation: Densely segmenting semantic objects and stuff at the pixel level (e.g., table, sofa, wall).
    3. Object Part Segmentation: Densely segmenting compositional components or functional parts of objects (e.g., chair leg, cushion, car wheel).
    4. Material Parsing: Pixel-level segmentation of physical surface materials (e.g., wood, plastic, metal, fabric).
    5. Texture Recognition: Identifying structural surface patterns and descriptors (e.g., knitted, stratified, marbled), which can be predicted at the image or localized pixel level.

  3. Knowl 3 — Heterogeneous Multi-Task Training Strategy for UPerNet

    algorithm

    To train UPerNet on heterogeneous datasets where images have annotations for only a subset of tasks and at varying granularities (pixel-wise vs. image-level), the training procedure alternates across data sources and applies task-specific losses and masking:

    Input: Training datasets D={Dscene,Dobject,Dpart,Dmaterial,Dtexture}D = \{D_{\text{scene}}, D_{\text{object}}, D_{\text{part}}, D_{\text{material}}, D_{\text{texture}}\}, initial learning rate η0=0.02\eta_0 = 0.02, maximum iterations T=100000T = 100000, power p=0.9p = 0.9
    Output: Trained UPerNet parameters Θ\Theta
    for iter=1\text{iter} = 1 to TT do
        Sample data source Dk∼DD_k \sim D with probability proportional to ∣Dk∣|D_k|
        Sample mini-batch Bk⊂DkB_k \subset D_k
        Update learning rate: η=η0⋅(1−iter/T)p\eta = \eta_0 \cdot (1 - \text{iter} / T)^p
        Forward propagate BkB_k through UPerNet
        if DkD_k corresponds to Scene then
            Compute cross-entropy loss on image-level scene logits
        else if DkD_k corresponds to Object or Material then
            Compute pixel-wise cross-entropy loss, masking out unlabeled/void pixels
        else if DkD_k corresponds to Part then
            Compute pixel-wise cross-entropy loss with an explicit background class, restricted strictly to the spatial region of the annotated parent object
        end if
        Backpropagate gradients and update only the network parameters along the active path for source DkD_k
    end for
    Freeze backbone layers
    for each texture fine-tuning epoch do
        Train texture convolutional head on cropped/resized patches using pixel-level predictions
    end for
    return Θ\Theta

    Training uses synchronized SGD across 8 GPUs (2 images per GPU, batch size 16) with synchronized batch normalization (channels = 512, momentum = 0.9, weight decay = 0.0001). Training image shorter sides are randomly scaled from {300,375,450,525,600}\{300, 375, 450, 525, 600\} with the maximum side restricted to 1200 pixels.

  4. Knowl 4 — Broden+ Dataset for Unified Perceptual Parsing

    data/table

    Broden+ is a standardized multi-source dataset created for training and evaluating models on Unified Perceptual Parsing. It consolidates annotations from ADE20K, Pascal-Context, Pascal-Part, OpenSurfaces, and the Describable Textures Dataset (DTD).

    Standardization involves:

    1. Merging overlapping object and part classes across ADE20K, Pascal-Context, and Pascal-Part.
    2. Filtering object classes to those appearing in ≥50\ge 50 images and ≥50,000\ge 50{,}000 pixels; filtering part classes to those in ≥20\ge 20 images.
    3. Merging rare or under-sampled material labels in OpenSurfaces (e.g., concrete and stone into stone) and removing classes with <50< 50 images.
    4. Mapping over 400 ADE20K scene classes to 365 Places categories.

    The dataset comprises 57,095 images total (51,617 training, 5,478 validation), divided across the following label types:

    Category Classes Sources Evaluation Metrics
    Scene 365 ADE Top-1 Accuracy
    Object 335 ADE, Pascal-Context mIoU (%) Pixel Accuracy (%)
    Object w/ part 77 ADE, Pascal-Context -
    Part 152 ADE, Pascal-Part mIoU-bg (%) Pixel Accuracy (%)
    Material 26 OpenSurfaces mIoU (%) Pixel Accuracy (%)
    Texture 47 DTD Top-1 Accuracy

    Note: For part evaluation, mIoU-bg measures intersection-over-union including background within object masks to prevent artificially inflated scores on small parts.

  5. Knowl 5 — ADE20K Semantic Segmentation Architecture Ablation

    data/table

    Ablation experiments on ADE20K semantic segmentation using ResNet-50 evaluate the components of UPerNet against state-of-the-art architectures without multi-scale testing:

    Method Mean IoU (%) Pixel Acc. (%) Overall (%) Training Time (hr)
    FCN 29.39 71.32 50.36 -
    SegNet 21.64 71.00 46.32 -
    DilatedNet 32.31 73.55 52.93 -
    CascadeNet 34.90 74.52 54.71 -
    RefineNet (Res-152) 40.70 - - -
    DilatedNet (Res-50) 34.28 76.35 55.32 53.9
    PSPNet (Res-50) 41.68 80.04 60.86 61.1
    FPN (/16) 34.46 76.04 55.25 18.1
    FPN (/8) 34.99 76.54 55.77 20.2
    FPN (/4) 35.26 76.52 55.89 21.2
    FPN + PPM (/4) 40.13 79.61 59.87 27.8
    FPN + PPM + Fusion (/4) 41.22 79.98 60.60 38.7

    Key observations:

    • FPN at stride 4 with PPM and multi-level feature fusion achieves 41.22% mIoU and 79.98% Pixel Accuracy, matching the performance of PSPNet (41.68% mIoU / 80.04% PA).
    • UPerNet reduces training time to 38.7 hours compared to PSPNet's 61.1 hours (requiring only ~63.3% of the training time for the same number of iterations), avoiding the heavy computational and memory burden of dilated convolutions across deep stages.

  6. Knowl 6 — Unified Perceptual Parsing Results on Broden+

    data/table

    Quantitative results on the Broden+ validation set evaluate the impact of progressively adding task supervision during multi-task training:

    Training Data Object Part Scene Material Texture
    +O +P +S +M +T mIoU P.A. mIoU(bg) P.A. Top-1 mIoU P.A. Top-1
    ✓ 24.72 78.03 - - - - - -
    ✓ - - - - - 52.78 84.32 -
    ✓ ✓ 23.92 77.48 30.21 48.30 - - - -
    ✓ ✓ ✓ 23.83 77.23 30.10 48.34 71.35 - - -
    ✓ ✓ ✓ ✓ 23.36 77.09 28.75 46.92 70.87 54.19 84.45 -
    ✓ ✓ ✓ ✓ ✓ 23.36 77.09 28.75 46.92 70.87 54.19 84.45 35.10

    Key observations:

    • Adding part and scene supervision causes minor drops in object segmentation (from 24.72% to 23.83% mIoU) while enabling competitive scene recognition (71.35% top-1 accuracy).
    • Joint multi-task training boosts material parsing performance from 52.78% mIoU (single-task baseline) to 54.19% mIoU (+1.41%), indicating that higher-level object and scene contextual features act as beneficial priors for material recognition.

  7. Knowl 7 — Visual Knowledge Graph Discovery from Natural Scenes

    model/method

    Hierarchical visual knowledge graphs are constructed by applying a trained UPerNet across large unannotated or scene-level datasets (such as the 36,500 validation images of Places-365) to extract co-occurrence and compositional relations:

    1. Scene-Object Relations: Modeled as a bipartite graph G=(V,E)G = (V, E) where node sets V=Vs∪VoV = V_s \cup V_o represent scenes vs∈Vsv_s \in V_s and objects vo∈Vov_o \in V_o. Weighted edges e(vs,vo)∈Ee(v_s, v_o) \in E represent the empirical percentage likelihood of object vov_o appearing in scene vsv_s, normalized by the frequency of scene vsv_s.
    2. Object-Material and Part-Material Relations: Extracted at the pixel level by aggregating pixels where an object class vov_o (or part class vpv_p) co-occurs with a material label vmv_m, determining the material composition distribution for each object or part.
    3. Material-Texture Relations: Constructed by linking pixel-level material predictions with texture classification predictions to identify visual surface descriptors characteristic of given materials (e.g., mapping carpet to matted, blotchy, crosshatched, or grooved textures).

Coverage note — None was omitted; all contributed models, task definitions, datasets, training strategies, ablation benchmarks, multi-task experiments, and visual knowledge discovery methods are fully covered.

References

  1. 1.He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016)
  2. 2.Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ADE20K dataset. In: Proceedings of CVPR (2017)
  3. 3.Hu, R., Dollar, P., He, K., Darrell, T., Girshick, R.: Learning to segment every thing. arXiv preprint arXiv:1711.10370 (2017)
  4. 4.Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3606–3613. IEEE (2014)
  5. 5.Liu, C., Sharan, L., Adelson, E.H., Rosenholtz, R.: Exploring features in a Bayesian framework for material recognition. In: Proceedings of CVPR (2010)
  6. 6.Bell, S., Upchurch, P., Snavely, N., Bala, K.: OpenSurfaces: a richly annotated catalog of surface appearance. ACM Trans. Graph. (TOG) 32(4), 111 (2013)
  7. 7.Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems, pp. 1097–1105 (2012)
  8. 8.Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2015)
  9. 9.Szegedy, C., et al.: Going deeper with convolutions. In: CVPR (2015)
  10. 10.Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Semantic image segmentation with deep convolutional nets and fully connected CRFs. In: International Conference on Learning Representations (ICLR) (2014)
  11. 11.Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3431–3440 (2015)
  12. 12.Noh, H., Hong, S., Han, B.: Learning deconvolution network for semantic segmentation. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 1520–1528 (2015)
  13. 13.Zeiler, M.D., Taylor, G.W., Fergus, R.: Adaptive deconvolutional networks for mid and high level feature learning. In: 2011 IEEE International Conference on Computer Vision (ICCV), pp. 2018–2025. IEEE (2011)
  14. 14.Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. In: International Conference on Learning Representations (ICLR) (2016)
  15. 15.Lin, G., Milan, A., Shen, C., Reid, I.: RefineNet: multi-path refinement networks for high-resolution semantic segmentation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)
  16. 16.Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2881–2890 (2017)
  17. 17.Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (VOC) challenge. Int. J. Comput. Vis. 88(2), 303–338 (2010)
  18. 18.Cordts, M., et al.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3213–3223 (2016)
  19. 19.Keeler, J.D., Rumelhart, D.E., Leow, W.K.: Integrated segmentation and recognition of hand-printed numerals. In: Advances in Neural Information Processing Systems, pp. 557–563 (1991)
  20. 20.Kokkinos, I., Maragos, P.: An expectation maximization approach to the synergy between image segmentation and object categorization. In: 2005 Tenth IEEE International Conference on Computer Vision, ICCV 2005, vol. 1, pp. 617–624. IEEE (2005)
  21. 21.Maire, M., Stella, X.Y., Perona, P.: Object detection and segmentation from joint embedding of parts and pixels. In: 2011 IEEE International Conference on Computer Vision (ICCV), pp. 2142–2149. IEEE (2011)
  22. 22.Elhoseiny, M., El-Gaaly, T., Bakry, A., Elgammal, A.: Convolutional models for joint object categorization and pose estimation. arXiv preprint arXiv:1511.05175 (2015)
  23. 23.Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 2650–2658 (2015)
  24. 24.Teichmann, M., Weber, M., Zoellner, M., Cipolla, R., Urtasun, R.: MultiNet: real-time joint semantic reasoning for autonomous driving. arXiv preprint arXiv:1612.07695 (2016)
  25. 25.Kokkinos, I.: UberNet: training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)
  26. 26.Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: quantifying interpretability of deep visual representations. In: Proceedings of CVPR (2017)
  27. 27.Mottaghi, R., et al.: The role of context for object detection and semantic segmentation in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)
  28. 28.Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: detecting and representing objects using holistic models and body parts. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)
  29. 29.Zhou, B., Bau, D., Oliva, A., Torralba, A.: Interpreting deep visual representations via network dissection. In: IEEE Trans. Pattern Anal. Mach. Intell. (2018)
  30. 30.Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., Oliva, A.: Learning deep features for scene recognition using places database. In: Advances in Neural Information Processing Systems, pp. 487–495 (2014)
  31. 31.Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: CVPR, vol. 1, p. 4 (2017)
  32. 32.Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Object detectors emerge in deep scene CNNs. In: International Conference on Learning Representations (ICLR) (2015)
  33. 33.Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2921–2929. IEEE (2016)
  34. 34.Ioffe, S., Szegedy, C.: Batch normalization: accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning, pp. 448–456 (2015)
  35. 35.Nair, V., Hinton, G.E.: Rectified linear units improve restricted Boltzmann machines. In: Proceedings of the 27th International Conference on Machine Learning (ICML 2010), pp. 807–814 (2010)
  36. 36.Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: DeepLab: semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. arXiv preprint arXiv:1606.00915 (2016)
  37. 37.Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: a large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2009, pp. 248–255. IEEE (2009)
  38. 38.Ioffe, S.: Batch renormalization: towards reducing minibatch dependence in batch-normalized models. In: Advances in Neural Information Processing Systems, pp. 1942–1950 (2017)
  39. 39.Peng, C., et al.: MegDet: a large mini-batch object detector. arXiv preprint arXiv:1711.07240 (2017)
  40. 40.Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5987–5995. IEEE (2017)
  41. 41.Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8689, pp. 818–833. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-10590-1_53
  42. 42.Badrinarayanan, V., Kendall, A., Cipolla, R.: SegNet: a deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 39(12), 2481–2495 (2017)
  43. 43.Kirillov, A., He, K., Girshick, R., Dollar, P.: Mscoco challenge 2017: stuff segmentation, team fair (2017)
  44. 44.Brandes, U., Robins, G., McCranie, A., Wasserman, S.: What is network science? Netw. Sci. 1(1), 1–15 (2013)

Citation

MLA
Xiao, T., et al. “Unified Perceptual Parsing for Scene Understanding”. Lecture Notes in Computer Science, Springer International Publishing, 2018, pp. 432–48, https://doi.org/10.1007/978-3-030-01228-1_26.
APA
Xiao, T., Liu, Y., Zhou, B., Jiang, Y., & Sun, J. (2018). Unified Perceptual Parsing for Scene Understanding. In Lecture Notes in Computer Science (pp. 432–448). Springer International Publishing. https://doi.org/10.1007/978-3-030-01228-1_26
Chicago
Xiao, T., Y. Liu, B. Zhou, Y. Jiang, and J. Sun. 2018. “Unified Perceptual Parsing for Scene Understanding”. In Lecture Notes in Computer Science. Springer International Publishing. https://doi.org/10.1007/978-3-030-01228-1_26.
Harvard
Xiao, T. et al. (2018) “Unified Perceptual Parsing for Scene Understanding”, Lecture Notes in Computer Science. Springer International Publishing, pp. 432–448. Available at: https://doi.org/10.1007/978-3-030-01228-1_26.
Vancouver
1. Xiao T, Liu Y, Zhou B, Jiang Y, Sun J (2018) Unified Perceptual Parsing for Scene Understanding. In: Lecture Notes in Computer Science. Springer International Publishing, pp 432–448

BibTeX

@inbook{Xiao_2018, title={Unified Perceptual Parsing for Scene Understanding}, ISBN={9783030012281}, ISSN={1611-3349}, url={http://dx.doi.org/10.1007/978-3-030-01228-1_26}, DOI={10.1007/978-3-030-01228-1_26}, booktitle={Computer Vision – ECCV 2018}, publisher={Springer International Publishing}, author={Xiao, Tete and Liu, Yingcheng and Zhou, Bolei and Jiang, Yuning and Sun, Jian}, year={2018}, pages={432–448} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF