LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop

Fisher YuAri SeffYinda ZhangShuran SongThomas FunkhouserJianxiong Xiao

article2015arXiv2,698 citations

Presents an efficient human-in-the-loop labeling framework that scales dataset construction using deep model predictions, producing the massive LSUN visual recognition benchmark spanning millions of scene and object images.

Listen

The article addresses the challenge of creating sufficiently large and dense labeled image datasets to train high-capacity deep learning models for visual recognition. Existing collections such as ImageNet and Places have become too small relative to modern network sizes, and fully manual labeling cannot scale fast enough to close the gap.

The work set out to build a new dataset, LSUN, containing roughly one million labeled images in each of ten scene categories and twenty object categories, while keeping human effort practical.

Researchers collected more than one hundred million candidate images per category through web search, then applied an iterative pipeline. In each round a modest number of images received manual labels from Amazon Mechanical Turk workers; those labels trained a convolutional classifier that confidently labeled easy positives and negatives and passed only ambiguous images to the next round. The process continued until the remaining set was small enough for exhaustive manual review. Quality was maintained through redundant labeling, tutorial checks, and hidden ground-truth images.

The resulting dataset contains ten million scene images and fifty-nine million object images. Average labeling precision reached approximately ninety percent. Human effort was amplified by a factor of roughly forty. When AlexNet was fine-tuned on the new data, classification error on a Places-derived test set fell from 28.6 percent to 22.2 percent. Pre-training the same architecture on LSUN object images raised mean average precision on PASCAL VOC 2012 from 0.80 to 0.87; a deeper VGG network showed a similar gain.

These outcomes indicate that training on larger but slightly noisy data can outperform training on smaller clean collections, and that first-layer filters learned from the denser set appear cleaner and more coherent. The gains matter because they demonstrate a practical route to continued progress in visual recognition without proportional increases in manual annotation cost.

The authors plan to keep expanding LSUN and to release it publicly. Practitioners should consider using the data for pre-training or fine-tuning when category coverage aligns with their needs, while monitoring for residual label noise on categories that contain many toys or rendered images. Additional validation on new tasks and architectures would strengthen before large-scale deployment.

The main limitations are the ninety-percent precision ceiling, possible dataset bias relative to other collections, and the focus on only two model families. Results should be treated as indicative rather than definitive for every downstream application.

Princeton University classification convolutional neural networks deep learning LSUN Large-Scale Image Dataset Deep Learning With Humans In The Loop

arXiv: 1506.03365 fyu/lsun

Paper: Learning Deep Features for Scene Recognition using Places Database, Bolei Zhou et al. (2014). It introduces the Places database and demonstrates how dedicated large-scale scene datasets enable convolutional networks to learn scene-centric features, directly motivating LSUN's creation and evaluation protocol.
Paper: ImageNet: A large-scale hierarchical image database, Jia Deng et al. (2009). It establishes the foundational methodology of using Amazon Mechanical Turk crowdsourcing to construct large-scale image ontologies that LSUN scales up through active human-in-the-loop filtering.
Paper: ImageNet Classification with Deep Convolutional Neural Networks, Alex Krizhevsky et al. (2012). It provides the baseline AlexNet architecture and pretraining methodology that LSUN fine-tunes to demonstrate the empirical benefits of large-scale scene data.
Paper: Unbiased look at dataset bias, A. Torralba et al. (2011). It formalizes the problem of dataset bias across visual recognition benchmarks, a key analytical limitation and motivation evaluated in the LSUN paper.
Paper: Microsoft COCO: Common Objects in Context, Tsung-Yi Lin et al. (2014). It provides the baseline pipeline for multi-stage crowdsourced visual data collection and standard object detection benchmarks used for downstream validation in LSUN.

Paper: Places: A 10 Million Image Database for Scene Recognition, Bolei Zhou et al. (2018). It scales dedicated scene-centric recognition to 10 million images across 434 categories, building directly on the large-scale scene learning paradigms explored by LSUN.
Paper: Revisiting Unreasonable Effectiveness of Data in Deep Learning Era, Chen Sun et al. (2017). It investigates whether scaling noisy, weakly labeled datasets by orders of magnitude continues to drive downstream transfer performance, extending LSUN's core findings.
Paper: LAION-5B: An open large-scale dataset for training next generation image-text models, Christoph Schuhmann et al. (2022). It massively scales automated web-scraping and filtering pipelines to billions of image-text pairs, extending the iterative, data-driven collection principles established in LSUN.
Paper: Semantic Understanding of Scenes Through the ADE20K Dataset, Bolei Zhou et al. (2016). It advances large-scale scene understanding from image-level category recognition to dense pixel-level parsing across complex visual environments.
Paper: Matterport3D: Learning from RGB-D Data in Indoor Environments, Angel Chang et al. (2017). It extends large-scale indoor scene dataset collection into full 3D spatial and RGB-D reconstructions for complex indoor visual reasoning.

Cover for LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop

Abstract

While there has been remarkable progress in the performance of visual recognition algorithms, the state-of-the-art models tend to be exceptionally data-hungry. Large labeled training datasets, expensive and tedious to produce, are required to optimize millions of parameters in deep network models. Lagging behind the growth in model capacity, the available datasets are quickly becoming outdated in terms of size and density. To circumvent this bottleneck, we propose to amplify human effort through a partially automated labeling scheme, leveraging deep learning with humans in the loop. Starting from a large set of candidate images for each category, we iteratively sample a subset, ask people to label them, classify the others with a trained model, split the set into positives, negatives, and unlabeled based on the classification confidence, and then iterate with the unlabeled set. To assess the effectiveness of this cascading procedure and enable further progress in visual recognition research, we construct a new image dataset, LSUN. It contains around one million labeled images for each of 10 scene categories and 20 object categories. We experiment with training popular convolutional networks and find that they achieve substantial performance gains when trained on this dataset.

Topics

classification convolutional neural networks deep learning labeling scheme recognition algorithm

Keywords

LSUN Large-Scale Image Dataset Deep Learning With Humans In The Loop Labeling Propagation Active Learning Amazon Mechanical Turk Dataset Density AlexNet VGG Pre-Training

1 Introduction
2 Overview
3 Data Collection
4 Deep Learning with Humans in the Loop
5 Crowd Sourcing
6 Results
7 Conclusion
References

Knowls

Knowl 1 — Cascaded Semi-Automated Dataset Annotation Pipeline
algorithm
The human-in-the-loop annotation pipeline expands candidate image sets into accurately labeled categories by iteratively learning a cascade of targeted, over-fit binary classifiers. In each round, human crowd annotators label a small subset of the remaining unlabeled pool. The classifier trained on accumulated human labels partitions the unlabeled dataset into confident positives (added to the target dataset), confident negatives (permanently discarded), and ambiguous images. Ambiguous images are passed to the next iteration along with newly sampled images, progressively concentrating human effort on difficult decision boundaries until the residual set is small enough for exhaustive manual annotation.
Input: Pool of unlabeled candidate images $U$ , target positive precision $p_{\text{pos}} = 0.95$ , target negative omission rate $r_{\text{neg}} = 0.01$ , sample batch size $N = 40000$
Output: Labeled positive set $P$ , discarded negative set $N_{\text{discard}}$
$P \leftarrow \emptyset$
$N_{\text{discard}} \leftarrow \emptyset$
$D_{\text{train}} \leftarrow \emptyset$
while $U$ is not small enough for exhaustive manual labeling do
    Sample $N$ images $S \subset U$ uniformly at random
    Partition $S$ into $S_{\text{test}}$ (10,000 images), $S_{\text{val}}$ (5,000 images), and $S_{\text{subtrain}}$ (25,000 images)
    Obtain doubly confirmed crowd labels for $S_{\text{test}}$ , $S_{\text{val}}$ , and $S_{\text{subtrain}}$ via AMT
     $D_{\text{train}} \leftarrow D_{\text{train}} \cup S_{\text{subtrain}}$
    Train classifier $f_\theta$ on $D_{\text{train}}$ and select best checkpoint on $S_{\text{val}}$
    Evaluate $f_\theta(x)$ for all $x \in S_{\text{test}}$
    Compute score threshold $\tau_{\text{pos}}$ such that $95\%$ of samples in $S_{\text{test}}$ with $f_\theta(x) \ge \tau_{\text{pos}}$ are true positives
    Compute score threshold $\tau_{\text{neg}}$ such that only $1\%$ of all true positives in $S_{\text{test}}$ have $f_\theta(x) < \tau_{\text{neg}}$
    for each $x \in U$ do
        if $f_\theta(x) \ge \tau_{\text{pos}}$ then
             $P \leftarrow P \cup \{x\}$
             $U \leftarrow U \setminus \{x\}$
        else if $f_\theta(x) < \tau_{\text{neg}}$ then
             $N_{\text{discard}} \leftarrow N_{\text{discard}} \cup \{x\}$
             $U \leftarrow U \setminus \{x\}$
        end if
    end for
    Retain ambiguous labeled samples from $S$ in $D_{\text{train}}$ for subsequent iterations
end while
Manually label all remaining images in $U$ and assign to $P$ or $N_{\text{discard}}$
return $P, N_{\text{discard}}$
Knowl 2 — Dual-Threshold Empirical Confidence Partitioning
model/method
To automatically partition candidate images while guaranteeing dataset quality, the classification score distribution of a trained model $f(x)$ is evaluated on an independently labeled validation test split $S_{\text{test}}$ . Two statistical thresholds, $\tau_{\text{pos}}$ and $\tau_{\text{neg}}$ , are empirically determined:
1. Positive Threshold ( $\tau_{\text{pos}}$ ): Defined as the classifier score value such that among all instances in $S_{\text{test}}$ scoring at or above $\tau_{\text{pos}}$ , $95\%$ are ground-truth positive annotations: $\frac{\sum_{x \in S_{\text{test}}} \mathbb{I}(f(x) \ge \tau_{\text{pos}} \land y_x = 1)}{\sum_{x \in S_{\text{test}}} \mathbb{I}(f(x) \ge \tau_{\text{pos}})} = 0.95$ where $y_x \in \{0, 1\}$ is the human ground-truth label for image $x$ .
2. Negative Threshold ( $\tau_{\text{neg}}$ ): Defined as the classifier score value such that no more than $1\%$ of the total true positive images in $S_{\text{test}}$ score strictly below $\tau_{\text{neg}}$ : $\frac{\sum_{x \in S_{\text{test}}} \mathbb{I}(f(x) < \tau_{\text{neg}} \land y_x = 1)}{\sum_{x \in S_{\text{test}}} \mathbb{I}(y_x = 1)} = 0.01$
Any image in the unlabeled pool scoring $f(x) \ge \tau_{\text{pos}}$ is automatically assigned a positive label and included in the dataset. Any image scoring $f(x) < \tau_{\text{neg}}$ is classified as negative and discarded. Images with scores $\tau_{\text{neg}} \le f(x) < \tau_{\text{pos}}$ are retained as the ambiguous subset and forwarded to subsequent labeling iterations.
Knowl 3 — Crowdsourced Verification Interface and Anti-Cheating Quality Control
model/method
Acquiring annotations from Amazon Mechanical Turk (AMT) relies on structured Human Intelligence Tasks (HITs) and multi-layered verification checks:
- HIT Structure: Each HIT comprises 205 fullscreen images displayed sequentially with category definitions, previous/next thumbnails, and binary toggle keybindings. Of these, 150 are candidate images and 55 are embedded control images.
- Redundant Doubly-Confirmed Labeling: Every image is independently labeled by two separate workers; only images with unanimous agreement are retained as ground truth.
- Tutorial and Gatekeeping (15 images): The initial 15 images of each HIT test comprehension of edge cases. Incorrect labels trigger modal pop-ups blocking progression until rectified.
- Online Ground Truth Check (20 images): Twenty embedded images with pre-verified labels are evaluated client-side. Workers failing to reach $\ge 90\%$ accuracy cannot submit the HIT and are prompted to review their labels.
- Hidden Ground Truth Check (20 images): Twenty ground truth images are validated server-side after submission without exposing their labels to the browser DOM, requiring $\ge 85\%$ accuracy to prevent script manipulation or selective online-check bypass.
Knowl 4 — Large-Scale Scene Understanding (LSUN) Dataset Structure
definition
The Large-Scale Scene Understanding (LSUN) dataset is a high-density visual benchmark designed to scale up the number of instances per visual concept. It contains labeled natural images across two domains:
- Scene Categories: 10 scene classes selected from the SUN database (bedroom, kitchen, living room, dining room, bridge, tower, restaurant, conference room, classroom, church outdoor), containing approximately 1 million labeled images per category (totaling 10 million scene images).
- Object Categories: 20 object classes matching PASCAL VOC 2012 (aeroplane, bicycle, bird, boat, bottle, bus, car, cat, chair, cow, dining table, dog, horse, motorbike, person, potted plant, sheep, sofa, train, tv/monitor), totaling 59 million labeled object images.
LSUN is approximately 10 times denser per category than the Places database and 100 times denser per basic category than ImageNet.
Knowl 5 — Search Query Generation via Adjective Augmentation and Time Slicing
model/method
To harvest over 100 million candidate image URLs per category while circumventing commercial search result quotas, web image queries (via Google Images) are expanded using two complementary techniques:
1. Adjective Augmentation: Category names are systematically combined with extensive lists of descriptive English adjectives (696 scene-relevant adjectives such as messy, sunny, desolate for scene categories, and dedicated object-relevant adjectives for object categories) to diversify visual appearances and broaden candidate retrieval.
2. Time Slicing: Each adjective-category query combination is restricted to 3-day publication intervals and queried exhaustively for all 3-day time spans from 2009 onward.
Harvested images undergo minimum-dimension filtering, retaining only images where $\min(\text{height}, \text{width}) > 256$ pixels (retaining roughly $60\%$ of downloads). Near-duplicate images with varied compression artifacts, crops, and resolutions are retained in the initial pool to serve as natural visual augmentations.

Knowl 6 — Transfer Classification Performance: LSUN vs. ImageNet Pre-Training on PASCAL VOC 2012

data/table

Pre-training deep convolutional networks on high-density category data (LSUN object categories sampled at 300,000 images per class across 20 classes) produces representations that generalize better to downstream classification tasks than pre-training on broader datasets with fewer examples per category (ImageNet with 1,000 classes). Models fine-tuned on the PASCAL VOC 2012 classification training set with Hinge loss and evaluated using 10-crop averaging on the PASCAL VOC 2012 validation set achieve consistently higher mean Average Precision (mAP) when pre-trained on LSUN.

Model	Pre-train	aero	bike	bird	boat	bottle	bus	car	cat	chair	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	tv	mAP
AlexNet	ImageNet	0.96	0.86	0.88	0.85	0.49	0.94	0.75	0.89	0.62	0.86	0.65	0.87	0.92	0.90	0.85	0.53	0.91	0.68	0.92	0.77	0.80
AlexNet	LSUN	0.98	0.93	0.94	0.90	0.64	0.95	0.78	0.97	0.74	0.96	0.72	0.96	0.98	0.94	0.85	0.59	0.96	0.76	0.97	0.82	0.87
VGG	ImageNet	0.95	0.85	0.92	0.86	0.65	0.92	0.80	0.93	0.68	0.77	0.66	0.90	0.85	0.86	0.94	0.56	0.89	0.60	0.94	0.85	0.82
VGG	LSUN	0.98	0.92	0.94	0.90	0.69	0.96	0.82	0.96	0.75	0.97	0.78	0.93	0.97	0.92	0.95	0.67	0.96	0.73	0.97	0.84	0.88

AlexNet pre-trained on LSUN demonstrates an overall mAP gain of $+0.07$ (0.87 vs. 0.80) compared to ImageNet pre-training. Similarly, VGG achieves $+0.06$ mAP improvement (0.88 vs. 0.82).

Knowl 7 — Scene Classification Accuracy Improvement on PLACES from Adding LSUN Data
empirical result
Augmenting scene classification training data with LSUN images significantly reduces classification error on the standard PLACES benchmark. Evaluating an AlexNet architecture on the 10 shared scene categories of the PLACES test set yields the following results:
- Fine-tuning AlexNet on PLACES alone results in an overall test classification error rate of $28.6\%$ .
- Fine-tuning AlexNet on PLACES combined with LSUN (capped at 200,000 images per category from LSUN) reduces the overall test error rate to $22.2\%$ .
- This corresponds to a $22.37\%$ relative reduction in testing error, with classification accuracy improving across 8 of the 10 individual scene categories (bedroom, kitchen, living room, dining room, bridge, tower, restaurant, and church outdoor).
Knowl 8 — Precision and Human Effort Amplification Rates in Semi-Automated Annotation
empirical result
Validation experiments evaluating the quality and efficiency of the semi-automated human-in-the-loop pipeline demonstrate:
- Label Precision: On an evaluation set of 2,000 images sampled across 11 object categories and labeled by trained expert annotators, average labeling precision is approximately $90\%$ . The primary sources of label noise are toy representations of objects, computer-rendered 3D models (especially within the person category), photo collages, and heavily edited imagery.
- Effort Amplification: The ratio of total images accurately categorized (assigned positive or negative) to the number of images explicitly labeled by human annotators averages $40:1$ across categories, exceeding $60:1$ for categories with distinctive visual features such as train.
Knowl 9 — First-Layer Convolutional Filter Quality under High Data Density
empirical result
Visual inspection of first-layer convolutional filters in AlexNet models reveals distinct qualitative differences depending on dataset density:
- Filters learned from scratch on ImageNet (broad coverage of 1,000 categories with ~1,000 images per class) exhibit substantial high-frequency noise, checkerboard artifacts, and less coherent edge orientations.
- Filters learned from scratch on LSUN object images (dense coverage of 20 categories with hundreds of thousands of images per class) converge to noticeably cleaner, smoother Gabor-like edge detectors and well-defined color opponent blobs.
Knowl 10 — Staged Classifier Architecture and Training Schedule in Label Propagation
experimental setup
The classification model used in the labeling cascade adapts across iterations to balance computational throughput against model capacity:
- Iterations 1–2: Features extracted from the final fully connected layer (fc7) of a pre-trained AlexNet are classified using a Multi-Layer Perceptron (MLP) with two hidden layers. This fast classifier eliminates over $50\%$ of easy negatives from the initial candidate pool.
- Iterations 3+: The pipeline transitions to fine-tuning a pre-trained GoogLeNet architecture using mini-batches of size 80. Fine-tuning uses an initial learning rate of $0.001$ , decaying to $0.0001$ after 40,000 iterations for a total of 60,000 training iterations.
- Inference Cropping Schedule: In early labeling iterations, inference is computed on single central crops to maximize speed across tens of millions of candidates. After iteration 10, when the unlabeled pool has been substantially reduced, inference evaluates 10 crops per image (center and four corners with horizontal mirroring) and averages the predictions.

Coverage note — None omitted; all primary contributions, including dataset statistics, querying strategies, cascaded labeling algorithms, crowdsourcing mechanisms, downstream transfer results, and filter representations, are fully documented.

References

1.http://www.image-net.org/challenges/LSVRC/announcement-June-2-2015.
2.S. Branson, C. Wah, F. Schroff, B. Babenko, P. Welinder, P. Perona, and S. Belongie. Visual recognition with humans in the loop. In ECCV, 2010.
3.B. Collins, J. Deng, K. Li, and L. Fei-Fei. Towards scalable dataset construction: An active learning approach. In ECCV, pages 86–98. Springer, 2008.
4.J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009.
5.J. Deng, O. Russakovsky, J. Krause, M. Bernstein, A. Berg, and L. Fei-Fei. Scalable multi-label annotation. In CHI, 2014.
6.L. Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. CVIU, 106(1):59–70, 2007.
7.K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. arXiv preprint arXiv:1502.01852, 2015.
8.S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
9.A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012.
10.Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998.
11.A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. arXiv preprint arXiv:1412.1897, 2014.
12.O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.
13.O. Russakovsky, L.-J. Li, and L. Fei-Fei. Best of both worlds: human-machine collaboration for object annotation. In CVPR, pages 2121–2131, 2015.
14.B. Settles. Active learning literature survey. University of Wisconsin, Madison, 52(55-66):11, 2010.
15.K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
16.C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. arXiv preprint arXiv:1409.4842, 2014.
17.C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.
18.S. Tong and D. Koller. Support vector machine active learning with applications to text classification. The Journal of Machine Learning Research, 2:45–66, 2002.
19.A. Torralba and A. A. Efros. Unbiased look at dataset bias. In CVPR, 2011.
20.S. Vijayanarasimhan and K. Grauman. Multi-level active prediction of useful image annotations for recognition. In D. Koller, D. Schuurmans, Y. Bengio, and L. Bottou, editors, Advances in Neural Information Processing Systems 21, pages 1705–1712. Curran Associates, Inc., 2009.
21.C. Wah, S. Branson, P. Perona, and S. Belongie. Multiclass recognition and part localization with humans in the loop. In ICCV, 2011.
22.R. Wu, S. Yan, Y. Shan, Q. Dang, and G. Sun. Deep image: Scaling up image recognition. arXiv preprint arXiv:1501.02876, 2015.
23.J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485–3492. IEEE, 2010.
24.B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. Learning deep features for scene recognition using places database. In NIPS, 2014.

Citation

MLA

Yu, F., et al. “LSUN: Construction of a Large-scale Image Dataset Using Deep Learning with Humans in the Loop”. arXiv, 2015, https://doi.org/10.48550/arxiv.1506.03365.

APA

Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., & Xiao, J. (2015). LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop. arXiv. https://doi.org/10.48550/arxiv.1506.03365

Chicago

Yu, F., A. Seff, Y. Zhang, S. Song, T. Funkhouser, and J. Xiao. 2015. “LSUN: Construction of a Large-scale Image Dataset Using Deep Learning with Humans in the Loop”. Preprint, ArXiv. https://doi.org/10.48550/arxiv.1506.03365.

Harvard

Yu, F. et al. (2015) “LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop”. arXiv. Available at: https://doi.org/10.48550/arxiv.1506.03365.

Vancouver

1. Yu F, Seff A, Zhang Y, Song S, Funkhouser T, Xiao J (2015) LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop. https://doi.org/10.48550/arxiv.1506.03365

BibTeX

@misc{https://doi.org/10.48550/arxiv.1506.03365,
  doi = {10.48550/ARXIV.1506.03365},
  url = {https://arxiv.org/abs/1506.03365},
  author = {Yu, Fisher and Seff, Ari and Zhang, Yinda and Song, Shuran and Funkhouser, Thomas and Xiao, Jianxiong},
  keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop},
  publisher = {arXiv},
  year = {2015},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

Metadata:DOI registry

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

License: Published with permission

LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Cascaded Semi-Automated Dataset Annotation Pipeline

Knowl 2 — Dual-Threshold Empirical Confidence Partitioning

Knowl 3 — Crowdsourced Verification Interface and Anti-Cheating Quality Control

Knowl 4 — Large-Scale Scene Understanding (LSUN) Dataset Structure

Knowl 5 — Search Query Generation via Adjective Augmentation and Time Slicing

Knowl 6 — Transfer Classification Performance: LSUN vs. ImageNet Pre-Training on PASCAL VOC 2012

Knowl 7 — Scene Classification Accuracy Improvement on PLACES from Adding LSUN Data

Knowl 8 — Precision and Human Effort Amplification Rates in Semi-Automated Annotation

Knowl 9 — First-Layer Convolutional Filter Quality under High Data Density

Knowl 10 — Staged Classifier Architecture and Training Schedule in Label Propagation

References

Citation

BibTeX

Source Code

Access the Paper