Back

Semantic Understanding of Scenes Through the ADE20K Dataset

Bolei ZhouHang ZhaoXavier PuigS. FidlerAdela BarriusoA. Torralba

article2016IJCV2,459 citations

Presents ADE20K, an open-vocabulary dataset of 25,000 densely annotated images spanning over 3,000 object and part classes, establishing standardized benchmarks and baselines for fine-grained scene parsing and instance segmentation.

Listen

Real-world autonomous systems and robotic platforms require comprehensive visual scene understanding to navigate spaces and interact with objects safely. While existing computer vision datasets often focus on high-level image classification or a restricted list of common objects, real environments contain dense, complex compositions of background regions, discrete objects, and functional object parts. High-quality, densely labeled visual data is critical to bridge this capability gap and allow intelligent systems to parse scenes at the pixel level.

The article introduces ADE20K, a densely annotated image dataset covering diverse scenes, objects, and hierarchical object parts, to benchmark and advance pixel-level computer vision models. It evaluates baseline and state-of-the-art architectures on scene parsing and instance segmentation tasks to demonstrate how detailed annotations improve visual recognition.

To construct ADE20K, a single expert annotator exhaustively segmented 25,210 complex everyday images across 900 scene categories using an open vocabulary, cataloging 3,169 distinct classes across objects, background stuff, and object parts. Benchmarks were established for scene parsing across the top 150 categories and instance segmentation across 100 foreground object classes. The authors then trained and evaluated various deep learning architectures, conducted ablation studies on technical training parameters such as batch normalization, and hosted competitive international challenge tracks to assess top-performing models against human-level baseline consistency.

The article yields several key findings. First, top semantic segmentation architectures achieve up to 81.01% pixel accuracy on the parsing benchmark, approaching the estimated human annotator consistency baseline of 82.4%. Second, incorporating multi-scale contextual features improves scene parsing mean intersection-over-union by 3% to 7% over standard baseline networks, and fusing instance segmentation output into scene parsing further boosts performance. Third, technical experiments reveal that synchronized batch normalization requires a minimum batch size of 4 to avoid significant performance degradation; dropping below this threshold reduces accuracy metrics by approximately 5%. Fourth, current top models still struggle with small, infrequent objects, heavy occlusions, reflective surfaces, and non-standard object orientations.

These findings demonstrate that high-density, multi-layer semantic annotations substantially advance practical machine vision, enabling practical downstream applications including automated object removal, scene synthesis, and hierarchical concept recognition. However, the persistent performance drop on small objects and non-canonical views indicates that current models still rely heavily on visual patterns rather than robust contextual reasoning, posing potential operational risks in cluttered or unfamiliar environments.

Organizations developing computer vision systems should adopt context-aware architectures and ensure training pipelines use synchronized batch normalization with adequate batch sizes across hardware processors. Further work is recommended to improve small-object detection, resolve reflective and occlusion ambiguities, and scale dense part-level annotations across larger training sets. Readers should note that open-vocabulary labeling introduces some naming and boundary ambiguity, with about 17.6% variance observed across repeated human annotations.

Massachusetts Institute of TechnologyPeking UniversityThe Chinese University of Hong KongUniversity of Torontoparsing algorithmscene understandingsegmentation performanceADE20KScene parsingInstance segmentation
arXiv: 1608.05442CSAILVision/semantic-segmentation-pytorch
Cover for Semantic Understanding of Scenes Through the ADE20K Dataset

Abstract

Semantic understanding of visual scenes is one of the holy grails of computer vision. Despite efforts of the community in data collection, there are still few image datasets covering a wide range of scenes and object categories with pixel-wise annotations for scene understanding. In this work, we present a densely annotated dataset ADE20K, which spans diverse annotations of scenes, objects, parts of objects, and in some cases even parts of parts. Totally there are 25k images of the complex everyday scenes containing a variety of objects in their natural spatial context. On average there are 19.5 instances and 10.5 object classes per image. Based on ADE20K, we construct benchmarks for scene parsing and instance segmentation. We provide baseline performances on both of the benchmarks and re-implement the state-of-the-art models for open source. We further evaluate the effect of synchronized batch normalization and find that a reasonably large batch size is crucial for the semantic segmentation performance. We show that the networks trained on ADE20K are able to segment a wide variety of scenes and objects1.

Topics

parsing algorithmscene understandingsegmentation performancereal scenes

Keywords

ADE20KScene parsingInstance segmentationSemantic segmentationBatch normalizationObject‑part relationshipPixel accuracyMean IoUMask R‑CNNUPerNet

Table of Contents

  • 1 Introduction
  • 1.1 Related work
  • 2 ADE20K: Fully Annotated Image Dataset
  • 2.1 Image annotation
  • 2.2 Dataset summary
  • 2.3 Annotation consistency
  • 2.4 Dataset statistics
  • 2.5 Object-part relationships
  • 2.6 Comparison with other datasets
  • 3 Pixel-wise Scene Understanding Benchmarks
  • 3.1 Scene parsing benchmark
  • 3.2 Opening source the state-of-the-art scene parsing models
  • 3.3 Effect of batch normalization for scene parsing
  • 3.4 Instance Segmentation
  • 3.5 How does scene parsing performance improve with instance information?
  • 4 Places Challenges
  • 4.1 Scene Parsing Track
  • 4.2 Instance Segmentation Track
  • 4.3 Take-aways from the Challenge
  • 5 Object-Part Joint Segmentation
  • 6 Applications
  • 7 Conclusion
  • References

Knowls

  1. Knowl 1 — ADE20K Dataset Properties and Multi-Level Annotation Taxonomy

    definition

    ADE20K is a densely annotated scene understanding dataset covering 900 scene categories. The dataset comprises 25,210 images partitioned into 20,210 training images, 2,000 validation images, and 3,000 testing images. All images were annotated by a single expert annotator using an open-ended vocabulary.

    Key structural characteristics and statistics of the dataset include:

    • Taxonomy and Classes: A total of 3,169 distinct semantic classes are labeled, comprising 2,693 object and stuff categories and 476 object part categories.
    • Annotation Density: Every image is exhaustively labeled with semantic concepts, containing at least 5 objects and up to 273 object instances (reaching 419 instances when object parts are included). On average, each image contains 19.5 object instances, 10.5 object classes, and 29 distinct polygon segments.
    • Object-Part Hierarchies: 76% of object instances include segmented part annotations (averaging 3 parts per object), forming part hierarchies up to depth 3 (for example, car →\rightarrow wheel →\rightarrow rim, or cabinet →\rightarrow door →\rightarrow knob). The category with the most part types is "building", with 79 distinct part classes. On average, 10% of image pixels are annotated as parts.
    • Instance Metadata: Each object instance includes metadata indicating whether it is occluded, cropped, or associated with specific visual attributes.

  2. Knowl 2 — Polygonal Annotation Depth Sorting for Dense Segmentation Mask Generation

    algorithm

    In datasets with dense, overlapping, and occluded polygon annotations, converting raw polygonal vector labels into a discrete 2D semantic segmentation mask requires depth ordering to prevent background or parent objects from improperly occluding foreground instances or object parts. The following procedure resolves polygon overlaps into a dense pixel mask:

    Input: Set of annotated polygons P={p1,p2,…,pN}P = \{p_1, p_2, \dots, p_N\} with class labels C={c1,c2,…,cN}C = \{c_1, c_2, \dots, c_N\} for an image of dimensions H×WH \times W
    Output: Dense semantic label mask M∈{0,…,K}H×WM \in \{0, \dots, K\}^{H \times W}
    Initialize relative depth layer d(pi)d(p_i) for all pi∈Pp_i \in P
    for each polygon pi∈Pp_i \in P do
        if cic_i is a background class (e.g., 'sky', 'wall') then
            Set d(pi)←∞d(p_i) \leftarrow \infty // Farthest depth layer
        end if
    end for
    for each pair of non-background polygons (pi,pj)(p_i, p_j) do
        if polygon pip_i is completely enclosed within polygon pjp_j then
            Set d(pi)<d(pj)d(p_i) < d(p_j) // Enclosed polygon is placed closer to viewer
        else if pip_i and pjp_j partially overlap then
            R←pi∩pjR \leftarrow p_i \cap p_j // Intersection region
            ni←n_i \leftarrow number of vertex points of pip_i lying inside RR
            nj←n_j \leftarrow number of vertex points of pjp_j lying inside RR
            if ni>njn_i > n_j then
                Set d(pi)<d(pj)d(p_i) < d(p_j) // Polygon with more vertices in intersection is closer
            else
                Set d(pj)<d(pi)d(p_j) < d(p_i)
            end if
        end if
    end for
    Sort polygons in PP in descending order of depth (farthest to closest)
    Initialize segmentation mask M(x,y)←unlabeledM(x, y) \leftarrow \text{unlabeled} for all (x,y)(x, y)
    for each polygon pp in sorted order from farthest to closest do
        for each pixel (x,y)(x, y) inside polygon pp do
            M(x,y)←label(p)M(x, y) \leftarrow \text{label}(p)
        end for
    end for
    return MM

    This depth ordering ensures that object parts never occlude whole objects and whole objects do not inappropriately occlude their constituent parts during rasterization.

  3. Knowl 3 — Annotation Consistency and Error Taxonomy in Open-Vocabulary Dense Segmentation

    empirical result

    In open-vocabulary, dense scene annotation, consistency analysis between repeated annotations reveals sources of labeling noise across time and annotators. In an experiment where 61 validation images were re-annotated by the same expert annotator after a 6-month interval, the average pixel agreement was 82.4% (median consistency 89.7%). The remaining 17.6% pixel discrepancies were categorized into three distinct error sources:

    1. Segmentation Quality (5.7% of pixels, median 4.8%): Boundary outlining variances and differences in precision when tracing complex or irregular object contours (such as trees or building boundaries).
    2. Object Naming Ambiguity (6.0% of pixels, median 0.3%): Inconsistencies in taxonomic naming due to concept hierarchy or semantic similarity (for example, labeling an entity as "car" versus "truck", or "palm tree" versus "tree").
    3. Segmentation Quantity / Missed Objects (5.9% of pixels, median 2.6%): Small or background objects segmented in one session but omitted in the other.

    When comparing the single expert annotator against two experienced external annotators on 20 validation images:

    • The first external annotator exhibited 58.5% pixel discrepancy relative to the expert (7.9% segmentation contour inconsistency, 14.9% naming inconsistency, and 35.8% missing object pixels).
    • The second external annotator exhibited 75.0% pixel discrepancy.
    • External annotators labeled an average of only 16 segments per image, compared to 29 segments per image produced by the expert annotator, demonstrating that single-expert annotation substantially reduces missing-object rates and naming divergence.

  4. Knowl 4 — SceneParse150 Benchmark Definition and Baseline Segmentation Performance

    data/table

    The SceneParse150 benchmark is built from ADE20K by selecting the top 150 categories ranked by total pixel coverage. The 150 categories comprise 35 stuff categories (such as wall, sky, and road) and 115 discrete object categories (such as car, person, and table). Together, these 150 classes account for 92.75% of all annotated pixels in ADE20K (60.92% stuff pixels and 31.83% discrete object pixels). Large images are rescaled so their minimum dimension is 512.

    The baseline semantic segmentation performance on the validation set across four standard metrics (Pixel Accuracy, Mean Accuracy, Mean Intersection-over-Union [mIoU], and Weighted IoU) is summarized below:

    Networks Pixel Acc. Mean Acc. Mean IoU Weighted IoU
    FCN-8s 71.32% 40.32% 0.2939 0.5733
    SegNet 71.00% 31.14% 0.2164 0.5384
    DilatedVGG 73.55% 44.59% 0.3231 0.6014
    DilatedResNet-34 76.47% 45.84% 0.3277 0.6068
    DilatedResNet-50 76.40% 45.93% 0.3385 0.6100
    Cascade-SegNet 71.83% 37.90% 0.2751 0.5805
    Cascade-DilatedVGG 74.52% 45.38% 0.3490 0.6108

    Dilated convolution backbones outperform standard downsampled encoder-decoder architectures (FCN-8s and SegNet). Adding cascade classification structures further improves performance, raising SegNet mIoU by ~5.9% and DilatedVGG mIoU by ~2.6%. Stuff classes (sky, road, building) achieve the highest IoU, while small or low-pixel-count object classes (blanket, tray, glass) achieve the lowest IoU.

  5. Knowl 5 — Batch Normalization Size and Synchronization Dynamics in Semantic Segmentation

    empirical result

    Semantic segmentation networks fine-tuned from ImageNet pre-trained backbones depend heavily on the effective Batch Normalization (BN) sample size rather than the aggregate mini-batch size. Using a Dilated ResNet-50 with a Pyramid Pooling Module (PSPNet) trained on ADE20K (8 GPUs, standard setting of 100k iterations and initial learning rate 0.02), varying BN synchronization modes and batch sizes yields the following validation performance:

    BN Status Batch Size BN Size Pixel Acc. Mean IoU
    Synchronized 16 16 79.73% 0.4126
    Synchronized 8 8 80.05% 0.4158
    Synchronized 4 4 79.71% 0.4119
    Synchronized 2 2 75.26% 0.3355
    Unsynchronized 16 2 75.28% 0.3403
    Frozen 16 N/A 78.32% 0.3809
    Frozen 8 N/A 78.29% 0.3793
    Frozen 4 N/A 78.34% 0.3833
    Frozen 2 N/A 78.81% 0.3856

    Key findings:

    • Critical Inflection Point: Under synchronized BN, performance remains steady for BN sizes from 16 down to 4 (mIoU ~0.412–0.416), but deteriorates drastically when BN size drops to 2 (mIoU drops by >7.6% to 0.3355). This establishes a critical BN sample size threshold of 4 for semantic segmentation, which is lower than the threshold of 16 observed in object detection. This difference arises because dense pixel-level supervision provides more statistical samples per image than sparse bounding boxes.
    • BN Size vs. Batch Size: In the unsynchronized setting with a total batch size of 16 (where each GPU computes BN statistics over only 2 images), the model yields 75.28% Pixel Acc and 0.3403 mIoU, matching the performance of a true batch size 2 model. Thus, BN sample size, not overall optimization batch size, determines representation quality.
    • Frozen BN Behavior: Freezing BN layers renders performance invariant to batch size (~78.3%–78.8% Pixel Acc, ~0.380–0.386 mIoU), but underperforms properly synchronized BN by over 3% mIoU.

  6. Knowl 6 — State-of-the-Art Multi-Scale Context Architectures on SceneParse150

    data/table

    Multi-scale contextual aggregation mechanisms significantly improve semantic scene parsing across both stuff and object categories on the SceneParse150 validation set. The table below compares plain dilated networks against the Pyramid Pooling Module (PPM) in PSPNet and the Feature Pyramid Network-style architecture in UPerNet:

    Networks Pixel Acc. Mean IoU
    DilatedResNet-18 77.41% 0.3534
    DilatedResNet-50 77.53% 0.3549
    DilatedResNet-18 + PPM 78.64% 0.3800
    DilatedResNet-50 + PPM 80.23% 0.4204
    DilatedResNet-101 + PPM 80.91% 0.4253
    UPerNet-50 80.23% 0.4155
    UPerNet-101 81.01% 0.4266

    Adding PPM to DilatedResNet-50 increases Mean IoU from 0.3549 to 0.4204 (+6.55%) and Pixel Accuracy from 77.53% to 80.23% (+2.70%). UPerNet-101 achieves 81.01% Pixel Accuracy and 0.4266 Mean IoU. These context aggregation modules enable the networks to capture scene layout and resolve local semantic ambiguities.

  7. Knowl 7 — InstSeg100 Instance Segmentation Benchmark and Baseline Mask R-CNN Evaluation

    data/table

    The InstSeg100 benchmark evaluates instance-level object segmentation on ADE20K, covering 100 foreground object categories with 218k total annotated instances (averaging 2.2k instances per class and 10 instances per image; all classes except "ship" have >100>100 instances).

    Baseline models based on Mask R-CNN with ResNet-50-FPN backbones evaluated on the InstSeg100 validation set yield the following performance across mean Average Precision (mAP\text{mAP}) and scale-stratified subsets (mAPS\text{mAP}_S for objects <32×32<32 \times 32 px, mAPM\text{mAP}_M for objects 32×3232 \times 32 to 96×9696 \times 96 px, and mAPL\text{mAP}_L for objects >96×96>96 \times 96 px):

    Networks mAPS\text{mAP}_S mAPM\text{mAP}_M mAPL\text{mAP}_L mAP\text{mAP}
    Mask R-CNN single-scale 0.0542 0.1737 0.2883 0.1832
    Mask R-CNN multi-scale 0.0733 0.2256 0.3584 0.2241

    Multi-scale training improves overall mAP by ~0.041. However, instance segmentation of small objects remains substantially more difficult (mAPS=0.0733\text{mAP}_S = 0.0733) than large objects (mAPL=0.3584\text{mAP}_L = 0.3584), with multi-scale training providing a +0.0701 gain on large objects but only a +0.0191 gain on small objects due to clutter and heavy occlusions in ADE20K scenes.

  8. Knowl 8 — Fusion of Instance Segmentation and Semantic Scene Parsing

    model/method

    Semantic scene parsing is inherently unaware of individual object instances, whereas instance segmentation models (such as Mask R-CNN) explicitly detect and delineate individual objects. To evaluate whether instance-level predictions improve scene parsing, outputs from Mask R-CNN (trained on 100 foreground classes) and PSPNet (trained on 150 classes) can be fused through late-stage post-processing:

    1. Filter Mask R-CNN predicted instance masks by confidence, retaining instances with confidence score ≥0.95\ge 0.95.
    2. Overlay these confident instance masks directly onto the dense semantic label map produced by PSPNet.
    3. For pixels where multiple instance masks overlap, assign the semantic class associated with the highest detection confidence. Pixels not covered by high-confidence instances retain the baseline PSPNet prediction (including all 35 stuff classes).

    Fusing Mask R-CNN predictions into PSPNet results in the following changes on the SceneParse150 validation set:

    • DilatedResNet-50 + PPM: Pixel Accuracy shifts from 80.23% to 80.21%, while Mean IoU increases from 0.4204 to 0.4256 (+0.52% mIoU).
    • DilatedResNet-101 + PPM: Pixel Accuracy remains at 80.91%, while Mean IoU increases from 0.4253 to 0.4290 (+0.37% mIoU).

    This demonstrates that object-instance boundaries and category predictions provide complementary spatial localization that enhances category-level scene parsing.

  9. Knowl 9 — Hierarchical Object-Part Joint Segmentation with Constrained Softmax Loss

    model/method

    ADE20K contains part annotations for 59 of its 150 standard object classes, encompassing 153 distinct part categories. Joint object and part segmentation can be formulated as a cascaded multi-task learning problem using UPerNet:

    • Training Formulation: The part segmentation branch predicts over part classes plus a background "non-part" class. During training, the softmax loss for part segmentation is computed conditionally: loss is calculated exclusively within the valid subset of part classes corresponding to the ground-truth object category of the pixel.
    • Inference Pipeline: During inference, the network first infers the primary object class mask. For each detected object region, part predictions are selected only from the specific candidate part set tied to that predicted object class.

    This hierarchical conditioning prevents impossible object-part combinations (such as predicting a "car wheel" inside a "person" region) and enables joint pixel-level parsing of objects and parts within a unified convolutional framework.

  10. Knowl 10 — Techniques and Performance in Places Challenge Instance Segmentation

    empirical result

    In the Places Challenge 2017 Instance Segmentation Track (evaluated on ADE20K InstSeg100 using COCO-style mAP metrics), top-performing architectures significantly outperformed the standard Mask R-CNN baseline:

    Team mAPS\text{mAP}_S mAPM\text{mAP}_M mAPL\text{mAP}_L mAP\text{mAP}
    Megvii (Face++) 0.1386 0.3015 0.4119 0.2977
    G-RMI 0.0980 0.2523 0.3858 0.2415
    Baseline Mask R-CNN 0.0733 0.2256 0.3584 0.2241

    The winning solution by Megvii (Face++) outperformed baseline Mask R-CNN by +0.0736 mAP overall and achieved a 41% relative improvement over G-RMI on small objects (mAPS=0.1386\text{mAP}_S = 0.1386 vs. 0.09800.0980). The primary methodological enhancements driving these results were:

    1. Precise RoI Pooling (PrRoIPooling): Treating discrete feature maps as continuous interpolated surfaces and computing two-order continuous integrals over proposal regions, eliminating discrete quantization and sample point approximations.
    2. Multi-Layer Proposal Context: Concatenating feature maps from preceding and succeeding feature pyramid layers for each proposal to inject surrounding spatial context necessary for resolving small and occluded objects.
    3. Bounding-Box Ensembling: Ensembling multiple detection bounding-box proposals prior to feeding them into the mask prediction branch.

Coverage note — Qualitative demonstrations of downstream applications (WordNet taxonomic grouping, automatic inpainting content removal, and generative scene synthesis) were omitted as standalone knowls because they apply existing external methods to ADE20K parsing maps rather than contributing new algorithmic or empirical findings.

References

  1. 1.Badrinarayanan V, Kendall A, Cipolla R (2017) Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans on Pattern Analysis and Machine Intelligence
  2. 2.Bell S, Upchurch P, Snavely N, Bala K (2013) OpenSurfaces: A richly annotated catalog of surface appearance. ACM Transactions on Graphics (TOG)
  3. 3.Bell S, Upchurch P, Snavely N, Bala K (2015) Material recognition in the wild with the materials in context database. In: Proc. CVPR
  4. 4.Caesar H, Uijlings J, Ferrari V (2017) Coco-stuff: Thing and stuff classes in context
  5. 5.Chen LC, Papandreou G, Kokkinos I, Murphy K, Yuille AL (2016) Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. arXiv:160600915
  6. 6.Chen X, Mottaghi R, Liu X, Cho NG, Fidler S, Urtasun R, Yuille A (2014) Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proc. CVPR
  7. 7.Cordts M, Omran M, Ramos S, Rehfeld T, Enzweiler M, Benenson R, Franke U, Roth S, Schiele B (2016) The cityscapes dataset for semantic urban scene understanding. In: Proc. CVPR
  8. 8.Dai J, He K, Sun J (2015) Convolutional feature masking for joint object and stuff segmentation. In: Proc. CVPR
  9. 9.Dai J, He K, Sun J (2016) Instance-aware semantic segmentation via multi-task network cascades. Proc CVPR
  10. 10.Everingham M, Van Gool L, Williams CK, Winn J, Zisserman A (2010) The pascal visual object classes (voc) challenge. Int’l Journal of Computer Vision
  11. 11.Geiger A, Lenz P, Urtasun R (2012) Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proc. CVPR
  12. 12.Goyal P, Dollár P, Girshick R, Noordhuis P, Wesolowski L, Kyrola A, Tulloch A, Jia Y, He K (2017) Accurate, large minibatch sgd: training imagenet in 1 hour. arXiv preprint arXiv:170602677
  13. 13.He K, Gkioxari G, Dollár P, Girshick R (2017) Mask r-cnn. In: Proc. ICCV
  14. 14.Huang JB, Kang SB, Ahuja N, Kopf J (2014) Image completion using planar structure guidance. ACM Transactions on Graphics (TOG)
  15. 15.Ioffe S, Szegedy C (2015) Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:150203167
  16. 16.Jiang B, Luo R, Mao J, Xiao T, Jiang Y (2018) Acquisition of localization confidence for accurate object detection. In: Proc. ECCV
  17. 17.Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks. In: In Advances in Neural Information Processing Systems
  18. 18.Lin TY, Maire M, Belongie S, Hays J, Perona P, Ramanan D, Dollár P, Zitnick CL (2014) Microsoft coco: Common objects in context. In: Proc. ECCV
  19. 19.Lin TY, Dollár P, Girshick R, He K, Hariharan B, Belongie S (2017) Feature pyramid networks for object detection. In: Proc. CVPR
  20. 20.Long J, Shelhamer E, Darrell T (2015) Fully convolutional networks for semantic segmentation. In: Proc. CVPR
  21. 21.Martin D, Fowlkes C, Tal D, Malik J (2001) A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proc. ICCV
  22. 22.Mottaghi R, Chen X, Liu X, Cho NG, Lee SW, Fidler S, Urtasun R, Yuille A (2014) The role of context for object detection and semantic segmentation in the wild. In: Proc. CVPR
  23. 23.Nathan Silberman PK Derek Hoiem, Fergus R (2012) Indoor segmentation and support inference from rgbd images. In: Proc. ECCV
  24. 24.Nguyen A, Dosovitskiy A, Yosinski J, Brox T, Clune J (2016) Synthesizing the preferred inputs for neurons in neural networks via deep generator networks
  25. 25.Noh H, Hong S, Han B (2015) Learning deconvolution network for semantic segmentation. In: Proc. ICCV
  26. 26.Patterson G, Hays J (2016) Coco attributes: Attributes for people, animals, and objects. In: Proc. ECCV
  27. 27.Peng C, Xiao T, Li Z, Jiang Y, Zhang X, Jia K, Yu G, Sun J (2018) Megdet: A large mini-batch object detector. In: Proc. CVPR, pp 6181–6189
  28. 28.Ren S, He K, Girshick R, Sun J (2015) Faster r-cnn: Towards real-time object detection with region proposal networks. In: In Advances in Neural Information Processing Systems
  29. 29.Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M, Berg AC, Fei-Fei L (2015) ImageNet Large Scale Visual Recognition Challenge. Int’l Journal of Computer Vision 115(3):211–252
  30. 30.Russell BC, Torralba A, Murphy KP, Freeman WT (2008) Labelme: a database and web-based tool for image annotation. Int’l Journal of Computer Vision
  31. 31.Song S, Lichtenberg SP, Xiao J (2015) Sun rgb-d: A rgb-d scene understanding benchmark suite. In: Proc. CVPR
  32. 32.Spain M, Perona P (2010) Measuring and predicting object importance. International Journal of Computer Vision
  33. 33.Wu Z, Shen C, van den Hengel A (2016) Wider or deeper: Revisiting the resnet model for visual recognition. CoRR abs/1611.10080, 1611.10080
  34. 34.Xiao J, Hays J, Ehinger KA, Oliva A, Torralba A (2010) Sun database: Large-scale scene recognition from abbey to zoo. In: Proc. CVPR
  35. 35.Xiao T, Liu Y, Zhou B, Jiang Y, Sun J (2018) Unified perceptual parsing for scene understanding. In: Proc. ECCV
  36. 36.Yu F, Koltun V (2016) Multi-scale context aggregation by dilated convolutions
  37. 37.Zhao H, Shi J, Qi X, Wang X, Jia J (2017) Pyramid scene parsing network. In: Proc. CVPR
  38. 38.Zhou B, Lapedriza A, Xiao J, Torralba A, Oliva A (2014) Learning deep features for scene recognition using places database. In: In Advances in Neural Information Processing Systems
  39. 39.Zhou B, Zhao H, Puig X, Fidler S, Barriuso A, Torralba A (2017) Scene parsing through ade20k dataset. In: Proc. CVPR

Citation

MLA
Zhou, B., et al. “Semantic Understanding of Scenes Through the ADE20K Dataset”. arXiv, 2016, http://arxiv.org/abs/1608.05442v2.
APA
Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., & Torralba, A. (2016). Semantic Understanding of Scenes through the ADE20K Dataset. arXiv. http://arxiv.org/abs/1608.05442v2
Chicago
Zhou, B., H. Zhao, X. Puig, et al. 2016. “Semantic Understanding of Scenes Through the ADE20K Dataset”. arXiv. http://arxiv.org/abs/1608.05442v2.
Harvard
Zhou, B. et al. (2016) “Semantic Understanding of Scenes through the ADE20K Dataset”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1608.05442v2.
Vancouver
1. Zhou B, Zhao H, Puig X, Xiao T, Fidler S, Barriuso A, Torralba A (2016) Semantic Understanding of Scenes through the ADE20K Dataset. arXiv

BibTeX

@article{zhou2016semantic,
  title = {Semantic Understanding of Scenes through the ADE20K Dataset},
  author = {Zhou, Bolei and Zhao, Hang and Puig, Xavier and Xiao, Tete and Fidler, Sanja and Barriuso, Adela and Torralba, Antonio},
  year = {2016},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1608.05442v2},
  eprint = {1608.05442}
}
Metadata:arXiv

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