LVIS: A Dataset for Large Vocabulary Instance Segmentation

Agrim GuptaPiotr DollárRoss B. Girshick

article2019CVPR1,874 citations

Introduces the LVIS benchmark spanning over 1,000 object categories and two million segmentation masks, exposing how poorly current deep learning models perform on long-tailed visual recognition with rare categories.

Listen

Modern computer vision algorithms perform exceptionally well when identifying a small set of common objects supported by thousands of examples. However, real-world visual environments naturally contain an extensive, long-tailed distribution of object categories, where many rare items appear infrequently. Existing benchmarks do not adequately measure how well systems recognize this vast variety of objects in low-sample settings, limiting the deployment of reliable image-understanding systems in complex environments.

The article introduces and evaluates LVIS, a large-scale instance segmentation benchmark designed to test whether machine learning algorithms can accurately delineate and categorize over 1,000 object classes, particularly in rare, low-data scenarios.

To make this evaluation feasible and fair, the authors designed a federated dataset structure and an iterative six-stage crowdsourcing pipeline across 164,000 images from the standard COCO collection. Instead of forcing every image to be exhaustively labeled for all 1,000+ categories—an impractical and ambiguous task—each category possesses a verified positive set (where every occurrence is masked) and a negative set (where the object does not appear). The dataset establishes high mask accuracy using human verification and WordNet synonym mapping rather than relying on automated labeling algorithms.

The investigation produced several key findings. First, existing top-tier architectures experience severe performance drops in rare-object regimes; for instance, standard models trained on reduced samples suffered an overall mask accuracy decline from 36.4% to 9.8%, with rare-category accuracy near zero (0.8%). Second, the proposed crowdsourced pipeline generated masks with higher boundary precision and overlap consistency relative to expert annotators than existing standards such as COCO and ADE20K. Third, simple algorithmic interventions yield immediate gains: adjusting inference detection thresholds and introducing a repeat factor sampling technique to oversample rare categories boosted rare-category average precision from 3.2% to 13.4% and overall precision to 27.1% when paired with larger backbones. Finally, empirical and simulation analyses revealed that evaluation set size inherently biases accuracy metrics downward on larger test sets due to the increased prevalence of rare categories.

These results demonstrate that extreme category imbalance and low-shot learning represent major bottlenecks for general-purpose visual recognition. Deploying current vision models in unconstrained environments introduces significant operational risk of missing uncommon objects. Furthermore, researchers must account for evaluation set size and category frequency shifts when measuring model performance and translating validation results to real-world test settings.

Organizations developing vision systems should integrate data resampling strategies, recalibrate inference-scoring thresholds, and benchmark models on large-vocabulary, long-tailed datasets before field deployment. For future research, the authors recommend participating in open benchmark challenges and refining architectures specifically to handle scarce training data.

The findings are bounded by the mid-stage release (v0.5), which represents approximately half of the projected two million masks and contains 1,230 categories across 82,000 images. Despite potential metric variance across differing evaluation set sizes, high experimental consistency across repeated runs and model ranks provides strong confidence in the benchmark's reliability.

Meta deep learning http object detection large vocabulary instance segmentation instance segmentation benchmark object detection dataset

arXiv: 1908.03195 lvis-dataset/lvis-api

Paper: Microsoft COCO: Common Objects in Context, Tsung-Yi Lin et al. (2014). Microsoft COCO provides the foundational benchmark and image base upon which LVIS is constructed to scale instance segmentation to a massive, long-tailed vocabulary.
Paper: Mask R-CNN, Kaiming He et al. (2017). Mask R-CNN establishes the primary baseline framework for instance segmentation that LVIS evaluates and benchmarks across high-cardinality, low-sample categories.
Paper: Semantic Understanding of Scenes Through the ADE20K Dataset, Bolei Zhou et al. (2016). ADE20K pioneered dense, large-vocabulary scene parsing and instance labeling, establishing essential precedents for open-vocabulary and multi-class visual recognition.
Paper: ImageNet Large Scale Visual Recognition Challenge, Olga Russakovsky et al. (2014). ImageNet introduced the 1,000-class scale to visual recognition that LVIS seeks to bring to pixel-level instance segmentation.
Paper: Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations, Ranjay Krishna et al. (2016). Visual Genome explores dense crowdsourced annotations and semantic relationship grounding across diverse object classes on COCO imagery.
Paper: DeViSE: A Deep Visual-Semantic Embedding Model, Andrea Frome et al. (2013). DeViSE introduces visual-semantic embedding methods that motivate zero-shot and open-vocabulary strategies needed to address LVIS's long-tailed object distributions.

Paper: Grounded Language-Image Pre-training, Liunian Harold Li et al. (2022). GLIP directly utilizes the LVIS benchmark to evaluate zero-shot and few-shot phrase grounding and large-vocabulary object detection.
Paper: Per-Pixel Classification is Not All You Need for Semantic Segmentation, Bowen Cheng et al. (2021). MaskFormer introduces a unified mask-classification framework specifically designed to scale efficiently to the large-vocabulary segmentation challenges highlighted by datasets like LVIS.
Paper: Segment Anything, Alexander M. Kirillov et al. (2023). Segment Anything builds on the need for exhaustive, high-quality segmentation across diverse vocabularies by introducing promptable foundation models trained at massive scale.
Paper: Intern VL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks, Zhe Chen et al. (2024). InternVL scales vision foundation models and leverages large-vocabulary benchmarks to assess transferability across complex visual-linguistic tasks.

Cover for LVIS: A Dataset for Large Vocabulary Instance Segmentation

Abstract

Progress on object detection is enabled by datasets that focus the research community's attention on open challenges. This process led us from simple images to complex scenes and from bounding boxes to segmentation masks. In this work, we introduce LVIS (pronounced `el-vis'): a new dataset for Large Vocabulary Instance Segmentation. We plan to collect ~2 million high-quality instance segmentation masks for over 1000 entry-level object categories in 164k images. Due to the Zipfian distribution of categories in natural images, LVIS naturally has a long tail of categories with few training samples. Given that state-of-the-art deep learning methods for object detection perform poorly in the low-sample regime, we believe that our dataset poses an important and exciting new scientific challenge. LVIS is available at this http URL.

Topics

deep learning http object detection training sample vocabulary segmentation methods

Keywords

large vocabulary instance segmentation instance segmentation benchmark object detection dataset LVIS Challenge COCO dataset exhaustive annotation segmentation masks entry-level object categories

1 Introduction
1.1 Related Datasets
2 Dataset Design
2.1 Task and Evaluation Overview
2.2 Federated Datasets
2.3 Evaluation Details
3 Dataset Construction
3.1 Annotation Pipeline
3.2 Vocabulary Construction
4 Dataset Analysis
4.1 Dataset Statistics
4.2 Annotation Consistency
4.3 Evaluation Protocol
5 Conclusion
A LVIS Release v0.5
B LVIS v0.5 Baselines
B.1 Mask R-CNN Out-of-the-Box
B.2 Mask R-CNN with Data Resampling
B.3 Mask R-CNN Standard Enhancements
C LVIS val to test Results Transfer
C.1 Category Frequency Distributions
C.2 AP as a Function of Evaluation Set Size
C.3 Comparing Models
C.4 Summary
References

Knowls

Knowl 1 — Federated Dataset Formulation for Object Instance Segmentation
definition
A federated dataset enables rigorous evaluation on a large vocabulary of categories $\mathcal{V}$ across a dataset of images $\mathcal{D}$ without requiring every image to be exhaustively annotated for all categories. For each category $c \in \mathcal{V}$ , the dataset defines two disjoint image subsets:
1. Positive set ( $P_c \subseteq \mathcal{D}$ ): Images in which instances of category $c$ appear and are exhaustively annotated.
2. Negative set ( $N_c \subseteq \mathcal{D}$ ): Images in which no instances of category $c$ appear.
The federated dataset is the union $\mathcal{D} = \bigcup_{c \in \mathcal{V}} (P_c \cup N_c)$ , where $|P_c \cup N_c| \ll |\mathcal{D}|$ for each $c$ .

Evaluation Protocol: During evaluation, an instance segmentation algorithm outputs predicted masks, class labels, and confidence scores across all test images without knowing the membership of images in $P_c$ or $N_c$ . The evaluation oracle computes Average Precision (AP) for category $c$ strictly over the evaluation subset $P_c \cup N_c$ . Predictions of category $c$ on images $i \notin P_c \cup N_c$ are ignored.

Non-Exhaustive Annotation Flag: Each image-category pair $(i, c)$ with $i \in P_c$ has an associated boolean flag $e_i^c$ . When $e_i^c = \text{True}$ (the default, $\sim 91\%$ of cases), annotations are exhaustive. When $e_i^c = \text{False}$ (e.g., in dense crowds where delineating every instance is impractical), false positive detections of category $c$ on image $i$ are not penalized during evaluation, though recall is measured against all labeled instances.
Knowl 2 — Repeat Factor Sampling for Class-Imbalanced Object Detection
model/method
Repeat Factor Sampling (RFS) is a training data resampling technique designed to mitigate extreme class imbalance in object detection and instance segmentation by oversampling images that contain rare categories.

Formulation: Let $f_c$ be the fraction of training images containing at least one instance of category $c$ . The category-level repeat factor $r_c$ is defined as:

$r_c = \max\left(1, \sqrt{\frac{t}{f_c}}\right)$

where $t$ is a frequency threshold hyperparameter. When $f_c \le t$ , the category is oversampled according to a square-root inverse frequency heuristic; when $f_c > t$ , $r_c = 1$ (no oversampling).

Because an image $i$ may contain multiple annotated categories, the image-level repeat factor $r_i$ is determined by the maximum repeat factor of the categories present in $i$ :

$r_i = \max_{c \in i} r_c$

where $\{c \in i\}$ is the set of categories annotated in image $i$ .

During training, SGD iterates over a random permutation of images where each image $i$ is replicated $\lfloor r_i \rfloor$ (or stochastically $\lceil r_i \rceil$ ) times per epoch.
Knowl 3 — Six-Stage Crowdsourced Annotation Pipeline
algorithm
The annotation pipeline collects high-quality instance segmentation masks over a large vocabulary without using machine learning algorithms in the data labeling loop.
Input: Image collection D, fixed vocabulary V
Output: Positive sets P_c with segmentation masks, exhaustive flags e_i^c, and negative sets N_c for all c in V
Stage 1 (Object Spotting):
    For each image i in D:
        Perform 2 independent iterative spotting passes where annotators mark one instance with a point and category c in V (using WordNet autocomplete).
        Retain category c for image i if and only if two annotators independently agree on the name.
        Yields initial positive sets P_c.
Stage 2 (Exhaustive Instance Marking):
    For each category c in V and image i in P_c:
        Present (i, c) to 5 annotators to verify the spotted instance against the definition of c and mark all other unspotted instances of c with points.
        Subsample frequent categories such that |P_c| <= 0.01 * |D|.
Stage 3 (Instance Segmentation):
    For each marked instance (i, o) from Stage 2:
        One annotator verifies the category label and draws a detailed polygon segmentation mask.
Stage 4 (Segment Verification):
    Iterate up to 4 rounds between Stage 3 and Stage 4:
        Present each polygon mask to up to 5 annotators.
        Accept the mask if >= 4 annotators rate it as high quality; otherwise requeue for Stage 3 re-annotation.
Stage 5 (Full Recall Verification):
    For each category c and image i in P_c:
        Ask up to 5 annotators if any unsegmented instances of c exist in image i.
        Set e_i^c = True if >= 4 annotators confirm all instances are segmented; set e_i^c = False if >= 2 report missing instances.
Stage 6 (Negative Set Collection):
    For each category c in V:
        Randomly sample images i in D \ P_c.
        Ask up to 5 annotators if category c appears in image i.
        If any annotator reports presence, reject image; otherwise add i to N_c.
        Sample until |N_c| = 0.01 * |D|.
Knowl 4 — Small Sample Evaluation Bias in Average Precision
theoretical result
In class-imbalanced evaluation settings, calculating Average Precision (AP) on smaller evaluation subsets introduces a systematic upward bias for rare categories compared to evaluation on larger datasets sampled from the identical distribution.

Under binary classification simulations with fixed positive frequency $f_c$ and class-conditional score distributions $p(s \mid y=1)$ and $p(s \mid y=0)$ (both for Gaussian score distributions and empirical neural network score distributions), estimated AP systematically decreases as the total evaluation size $n$ (or positive count $|P_c|$ ) increases, until reaching convergence. This decay with dataset size is most pronounced for low-frequency classes ( $f_c \ll 1$ ).

In real object detection, this phenomenon causes AP on a small validation set (e.g., 5k images) to be substantially higher than on a larger test set (e.g., 20k images) drawn from the same distribution when the dataset contains a large tail of rare categories.
Knowl 5 — LVIS Dataset Vocabulary and Frequency Bins
definition
LVIS (Large Vocabulary Instance Segmentation) defines categories using WordNet synsets to resolve synonyms and provide explicit visual definitions. The vocabulary $\mathcal{V}$ is constructed by winnowing an initial set of 8.8k concrete noun synsets down to 1,723 synsets via iterative spotting on COCO images.

Categories are partitioned into three disaggregated frequency bins based on the number of training images in which they appear:
1. Rare (r): Categories appearing in 1 to 10 training images.
2. Common (c): Categories appearing in 11 to 100 training images.
3. Frequent (f): Categories appearing in $>100$ training images.
In LVIS release v0.5, the dataset contains 1,230 categories across 82k images (57k train, 5k val, and 20k test), comprising 694k training instances and 51k validation instances. Approximately $75\%$ of categories appear in 100 or fewer training images.

Knowl 6 — Mask R-CNN Baseline and Resampling Performance on LVIS v0.5

data/table

Mask R-CNN (ResNet-50-FPN) trained out-of-the-box on LVIS v0.5 achieves near-zero mask Average Precision on rare categories ( $\text{AP}_r = 0.8\%$ ). Adjusting inference parameters (increasing maximum detections per image from 100 to 300 and lowering the minimum score threshold from 0.05 to 0.00) improves overall AP from $14.8\%$ to $21.0\%$ . Repeat Factor Sampling (RFS) with $t = 0.001$ further raises $\text{AP}_r$ to $13.4\%$ , and combining RFS with multiscale jitter and a ResNeXt-101-32x8d backbone achieves an overall AP of $27.1\%$ .

Configuration	$\text{AP}$	$\text{AP}_r$	$\text{AP}_c$	$\text{AP}_f$
Default inference (thr=0.05, det=100)	$14.8 \pm 0.19$	$0.6 \pm 0.21$	$11.0 \pm 0.36$	$25.2 \pm 0.10$
Tuned inference (thr=0.00, det=300)	$21.0 \pm 0.17$	$3.2 \pm 0.35$	$21.3 \pm 0.45$	$27.7 \pm 0.12$
+ RFS ( $t=0.0001$ )	$21.2 \pm 0.14$	$4.5 \pm 0.47$	$21.5 \pm 0.37$	$27.6 \pm 0.14$
+ RFS ( $t=0.0010$ )	$23.2 \pm 0.21$	$13.4 \pm 0.80$	$23.2 \pm 0.32$	$27.1 \pm 0.07$
+ RFS ( $t=0.0100$ )	$21.8 \pm 0.25$	$9.8 \pm 1.27$	$22.7 \pm 0.48$	$25.6 \pm 0.13$
Class-Aware Sampling (CAS)	$18.7 \pm 0.46$	$8.5 \pm 1.56$	$19.0 \pm 0.45$	$22.3 \pm 0.19$
RFS ( $t=0.001$ ) + Scale Jitter	$24.4 \pm 0.06$	$14.5 \pm 0.67$	$24.3 \pm 0.37$	$28.4 \pm 0.12$
+ ResNet-101-FPN	$26.0 \pm 0.18$	$15.8 \pm 0.95$	$26.1 \pm 0.21$	$29.8 \pm 0.22$
+ ResNeXt-101-32 $\times$ 8d-FPN	$27.1 \pm 0.43$	$15.6 \pm 1.14$	$27.5 \pm 0.77$	$31.4 \pm 0.12$

Knowl 7 — Segmentation Mask Quality and Boundary Complexity Comparison

data/table

When evaluated against independent expert annotators on 100 randomly sampled object instances ( $>32^2$ pixels), LVIS segmentation masks exhibit higher mask Intersection over Union (IoU) and boundary quality ( $F$ -measure) than COCO and ADE20K, while simultaneously possessing higher boundary complexity (measured as perimeter divided by $\sqrt{\text{area}}$ ). Confidence intervals are $95\%$ bootstrapped intervals.

Dataset	Mask IoU		Boundary Quality ( $F$ )		Boundary Complexity
	Mean	Median	Mean	Median	Mean	Median
COCO dataset vs. experts	0.83–0.87	0.88–0.91	0.77–0.82	0.79–0.88	5.59–6.04	5.13–5.51
COCO expert1 vs. expert2	0.91–0.95	0.96–0.98	0.92–0.96	0.97–0.99	6.94–7.84	5.86–6.80
ADE20K dataset vs. experts	0.84–0.88	0.90–0.93	0.83–0.87	0.84–0.92	6.00–6.84	4.79–5.31
ADE20K expert1 vs. expert2	0.90–0.94	0.95–0.97	0.90–0.95	0.99–1.00	6.34–7.43	4.83–5.53
LVIS dataset vs. experts	0.90–0.92	0.94–0.96	0.87–0.91	0.93–0.98	6.35–7.07	5.44–6.00
LVIS expert1 vs. expert2	0.93–0.96	0.96–0.98	0.91–0.96	0.97–1.00	7.13–8.48	5.91–6.82

Knowl 8 — Fidelity of Federated Dataset AP Evaluation Relative to Dense Annotations

data/table

To verify that sparse federated evaluation reproduces the evaluation accuracy of dense ground truth annotations, Mask R-CNN models trained on COCO were evaluated on the 80 COCO categories using both dense COCO validation annotations and sparse LVIS annotations on the same 5,000 images.

Model	Test Annotation Set	Box AP	Mask AP
ResNet-50-FPN	COCO	38.2	34.1
ResNet-50-FPN	LVIS	38.8	34.4
ResNet-101-FPN	COCO	40.6	36.0
ResNet-101-FPN	LVIS	40.9	36.0
ResNeXt-101-64 $\times$ 4d-FPN	COCO	47.8	41.2
ResNeXt-101-64 $\times$ 4d-FPN	LVIS	48.6	41.7

The close alignment across all models confirms that sparse federated evaluation is an accurate proxy for dense ground-truth benchmarking.

Coverage note — None was omitted; all key contributions including federated dataset formulation, annotation and vocabulary pipelines, repeat factor sampling, empirical baselines, evaluation bias analyses, and dataset quality evaluations are included.

References

1.Fred Attneave and Malcolm D Arnoult. The quantitative study of shape and pattern perception. Psychological bulletin, 1956. 7
2.Léon Bottou. Stochastic gradient descent tricks. In Neural networks: Tricks of the trade, pages 421–436. Springer, 2012. 10
3.Marc Brysbaert, Amy Beth Warriner, and Victor Kuperman. Concreteness ratings for 40 thousand generally known english word lemmas. Behavior research methods, 2014. 6
4.Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2
5.Piotr Dollár, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: An evaluation of the state of the art. TPAMI, 2012. 2
6.Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The PASCAL Visual Object Classes (VOC) Challenge. IJCV, 2010. 2
7.Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. TPAMI, 2006. 2
8.Ross Girshick, Ilija Radosavovic, Georgia Gkioxari, Piotr Dollár, and Kaiming He. Detectron. https://github.com/facebookresearch/detectron, 2018. 8
9.Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 9
10.Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In ICCV, 2017. 8
11.Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 2009. 10
12.Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 8, 9
13.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2
14.Derek Hoiem, Yodsawalai Chodpathumwan, and Qieyun Dai. Diagnosing error in object detectors. In ECCV. 2012. 11
15.Sergey Ioffe and Christian Szegedy. Batch normalization: accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2
16.Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollár. Panoptic segmentation. In CVPR, 2019. 7
17.Alex Krizhevsky, Ilya Sutskever, and Geoff Hinton. ImageNet classification with deep convolutional neural networks. In NIPS, 2012. 2
18.Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Tom Duerig, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. arXiv preprint arXiv:1811.00982, 2018. 3
19.Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation, 1989. 2
20.Yann LeCun, Corinna Cortes, and Christopher J.C. Burges. The MNIST database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 2
21.Marc Liberman. Reproducible research and the common task method. Simmons Foundation Lecture https://www.simonsfoundation.org/lecture/reproducible-research-and-the-common-task-method/, 2015. 2
22.Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 9
23.Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2
24.Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. COCO detection evaluation. http://cocodataset.org/#detection-eval, Accessed Oct 30, 2018. 2, 3
25.Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens van der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 10
26.David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 2, 7
27.Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. 10
28.George Miller. WordNet: An electronic lexical database. MIT press, 1998. 4
29.Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 2
30.Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 2
31.Bryan C Russell, Antonio Torralba, Kevin P Murphy, and William T Freeman. Labelme: a database and web-based tool for image annotation. IJCV, 2008. 1
32.Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In ECCV, 2016. 10
33.Merrielle Spain and Pietro Perona. Measuring and predicting importance of objects in our visual world. Technical Report CNS-TR-2007-002, California Institute of Technology, 2007. 1
34.Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The iNaturalist species classification and detection dataset. In CVPR, 2018. 2
35.Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 1
36.Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 10
37.Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ADE20K dataset. IJCV, 2019. 2
38.George Kingsley Zipf. The psycho-biology of language: An introduction to dynamic philology. Routledge, 2013. 1

Citation

MLA

Gupta, A., et al. “LVIS: A Dataset for Large Vocabulary Instance Segmentation”. arXiv, 2019, http://arxiv.org/abs/1908.03195v2.

APA

Gupta, A., Dollár, P., & Girshick, R. (2019). LVIS: A Dataset for Large Vocabulary Instance Segmentation. arXiv. http://arxiv.org/abs/1908.03195v2

Chicago

Gupta, A., P. Dollár, and R. Girshick. 2019. “LVIS: A Dataset for Large Vocabulary Instance Segmentation”. arXiv. http://arxiv.org/abs/1908.03195v2.

Harvard

Gupta, A., Dollár, P. and Girshick, R. (2019) “LVIS: A Dataset for Large Vocabulary Instance Segmentation”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1908.03195v2.

Vancouver

1. Gupta A, Dollár P, Girshick R (2019) LVIS: A Dataset for Large Vocabulary Instance Segmentation. arXiv

BibTeX

@article{gupta2019lvis,
  title = {LVIS: A Dataset for Large Vocabulary Instance Segmentation},
  author = {Gupta, Agrim and Dollár, Piotr and Girshick, Ross},
  year = {2019},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1908.03195v2},
  eprint = {1908.03195}
}

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

License: IEEE

LVIS: A Dataset for Large Vocabulary Instance Segmentation

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Federated Dataset Formulation for Object Instance Segmentation

Knowl 2 — Repeat Factor Sampling for Class-Imbalanced Object Detection

Knowl 3 — Six-Stage Crowdsourced Annotation Pipeline

Knowl 4 — Small Sample Evaluation Bias in Average Precision

Knowl 5 — LVIS Dataset Vocabulary and Frequency Bins

Knowl 6 — Mask R-CNN Baseline and Resampling Performance on LVIS v0.5

Knowl 7 — Segmentation Mask Quality and Boundary Complexity Comparison

Knowl 8 — Fidelity of Federated Dataset AP Evaluation Relative to Dense Annotations

References

Citation

BibTeX

Source Code

Access the Paper