Learning Deep Features for Scene Recognition using Places Database

Bolei ZhouÀgata LapedrizaJianxiong XiaoA. TorralbaA. Oliva

article2014NeurIPS3,103 citations

Introduces the 7-million-image Places database and demonstrates that convolutional neural networks trained on scene-centric data develop distinct visual representations that significantly outperform object-trained models on scene recognition tasks.

Listen

Scene recognition has long trailed object recognition in computer vision because available scene datasets were too small to train modern deep networks effectively. Researchers addressed this gap by building Places, a scene-centric collection of more than seven million labeled images across 476 categories, and by training convolutional neural networks on it to learn dedicated scene features.

The work set out to test whether a large, dedicated scene database could produce deep features that outperform those derived from the object-centric ImageNet dataset on scene classification tasks. The team gathered images through web searches that combined scene names with descriptive adjectives, removed duplicates, and used two rounds of Amazon Mechanical Turk verification to ensure label quality. They then trained a standard convolutional network architecture on a 205-category subset of Places and compared its performance and internal representations against an identical network trained on ImageNet. Relative density and diversity were measured through controlled human similarity judgments on the 88 categories shared with ImageNet and SUN.

Places proved comparable in density yet substantially more diverse than prior scene collections. The resulting Places-trained network reached 50 percent top-1 accuracy on the Places 205 test set and 66 percent on SUN 205, well above the accuracy obtained with ImageNet features. On established scene benchmarks it set new records, including 54.3 percent on SUN397 and 68.2 percent on MIT Indoor67, while ImageNet features remained stronger on object-centric tasks such as Caltech101. Visualization of network units showed that scene training produced receptive fields dominated by spatial layout and landscape structure, whereas object training produced more localized object-like patterns.

These outcomes indicate that scene and object recognition benefit from distinct high-level features and that large-scale scene data can close the performance gap that previously existed between the two domains. The two networks therefore offer complementary strengths that can be combined for broader visual recognition systems.

Further gains are likely from hybrid training sets that merge Places and ImageNet, from fine-tuning on additional scene benchmarks, and from relating the learned representations to known scene-processing pathways in the human brain. Results rest on the 205- and 476-category subsets released so far; performance on rarer or fine-grained scene distinctions remains untested, and all models inherit the biases present in web-sourced imagery.

Massachusetts Institute of Technology Princeton University Universitat Oberta de Catalunya database systems Places Database Convolutional Neural Network Scene Recognition

CSAILVision/places365 Zhou et al (2014).pdf

Paper: ImageNet Classification with Deep Convolutional Neural Networks, Alex Krizhevsky et al. (2012). This landmark paper introduced the deep convolutional neural network architectures and training protocols that the source study adapts and scales for large-scale scene recognition.
Paper: Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories, Svetlana Lazebnik et al. (2006). This foundational work establishes spatial pyramid matching for scene categorization, providing critical context on pre-deep-learning representations that the source study supersedes with CNN features.
Paper: Modeling the Shape of the Scene: A Holistic Representation of the Spatial Envelope, Aude Oliva et al. (2001). This classical study on holistic spatial properties of scenes defines the core perceptual dimensions of environments that the source study's deep network learns to represent implicitly.
Paper: ImageNet: A large-scale hierarchical image database, Jia Deng et al. (2009). This foundational paper introduces the ImageNet database, whose object-centric design and hierarchical structure motivated the creation of the scene-centric Places database presented in the source study.

Paper: Places: A 10 Million Image Database for Scene Recognition, Bolei Zhou et al. (2018). This later work directly extends the source study by scaling the Places database to over 10 million images and establishing comprehensive benchmark evaluations across multiple architectures.
Paper: NetVLAD: CNN Architecture for Weakly Supervised Place Recognition, Relja Arandjelović et al. (2015). This study builds upon the feature representations established by scene- and object-centric networks to develop an end-to-end differentiable architecture for visual place recognition.
Paper: Learning Deep Features for Discriminative Localization, Bolei Zhou et al. (2016). This subsequent work builds on the convolutional feature maps analyzed in the source study to introduce class activation mapping for weakly supervised localization.
Paper: Scene Parsing through ADE20K Dataset, Bolei Zhou et al. (2017). This later paper extends the scene-centric data collection paradigm pioneered in the source study to build ADE20K for exhaustive dense scene parsing.

Cover for Learning Deep Features for Scene Recognition using Places Database

Abstract

Scene recognition is one of the hallmark tasks of computer vision, allowing definition of a context for object recognition. Whereas the tremendous recent progress in object recognition tasks is due to the availability of large datasets like ImageNet and the rise of Convolutional Neural Networks (CNNs) for learning high-level features, performance at scene recognition has not attained the same level of success. This may be because current deep features trained from ImageNet are not competitive enough for such tasks. Here, we introduce a new scene-centric database called Places with over 7 million labeled pictures of scenes. We propose new methods to compare the density and diversity of image datasets and show that Places is as dense as other scene datasets and has more diversity. Using CNN, we learn deep features for scene recognition tasks, and establish new state-of-the-art results on several scene-centric datasets. A visualization of the CNN layers' responses allows us to show differences in the internal representations of object-centric and scene-centric networks.

Topics

database systems

Keywords

Places Database Convolutional Neural Network Scene Recognition Density Diversity ImageNet SUN Database Deep Feature Linear SVM Hybrid CNN

1 Introduction
2 Places Database
2.1 Building the Places Database
3 Comparing Scene-centric Databases
3.1 Relative Density and Diversity
3.2 Experimental Results
3.3 Cross Dataset Generalization
4 Training Neural Network for Scene Recognition and Deep Features
4.1 Visualization of the Deep Features
4.2 Results on Places 205 and SUN 205
4.3 Generic Deep Features for Visual Recognition
5 Conclusion
References

Knowls

Knowl 1 — Places Database Construction and Subsets
model/method
The Places database is a large-scale scene-centric image dataset containing 7,076,580 annotated images across 476 scene categories, inheriting its scene category taxonomy from the SUN database.

The database curation pipeline consists of three primary stages:
1. Query Formulation and Image Harvesting: 696 common English adjectives (e.g., messy, sunny, desolate) are combined with each of the 476 category names to form queries submitted to Google Images, Bing Images, and Flickr, yielding over 40 million downloaded images.
2. Filtering and De-duplication: Only color images with resolutions of $200 \times 200$ pixels or higher are retained. Exact duplicate URLs are eliminated, and PCA-based duplicate removal is executed within each category and across identical categories in the SUN database to prevent dataset overlap.
3. Crowdsourced Verification: Images undergo a two-round verification process on Amazon Mechanical Turk (AMT):
  - In the first round, workers view candidate images one at a time with category definitions; the default response is set to "No", requiring active selection of positive scene exemplars.
  - In the second round, positive images from the first round are re-annotated with the default response set to "Yes" to prune remaining false positives.
  - Quality control is enforced by randomly injecting 30 ground-truth positive and 30 ground-truth negative images from the SUN database into each Human Intelligence Task (HIT) of 750 images; assignments with control accuracy below 90% are discarded.
Two standardized benchmark subsets are defined:
- Places 205: The subset of 205 categories containing at least 5,000 images each. It includes 2,448,873 training images (between 5,000 and 15,000 images per category), a validation set of 100 images per category (20,500 total), and a test set of 200 images per category (41,000 total).
- Places 88: The subset of 88 categories that overlap with ImageNet and possess at least 1,000 images in ImageNet.
Knowl 2 — Formulations for Relative Dataset Density and Diversity
equation
To quantitatively compare image datasets without defining absolute visual distance scales, relative density and relative diversity are formulated using rank-ordered visual distances between image pairs.

Let $d(x, y)$ denote a visual distance metric between two images $x$ and $y$ . For two datasets $A$ and $B$ , let $a_1, a_2 \in A$ and $b_1, b_2 \in B$ be samples from each respective dataset.

Relative Density: Measures data concentration based on the proximity of nearest neighbors. If $a_2$ and $b_2$ are the nearest neighbors of $a_1$ and $b_1$ in their respective datasets, the relative density of $A$ with respect to $B$ is: $\text{Den}_B(A) = p(d(a_1, a_2) < d(b_1, b_2))$ Dataset $A$ is denser than $B$ if and only if $\text{Den}_B(A) > \text{Den}_A(B)$ . For $N$ datasets $A_1, A_2, \dots, A_N$ , the generalized relative density of $A_1$ is: $\text{Den}_{A_2,\dots,A_N}(A_1) = p\left(d(a_{11}, a_{12}) < \min_{i=2}^N d(a_{i1}, a_{i2})\right)$ where $a_{i1} \in A_i$ is a randomly selected image and $a_{i2} \in A_i$ is a near neighbor of $a_{i1}$ .

Relative Diversity: Inspired by the Simpson index of ecological diversity, relative diversity measures the probability that two randomly drawn images from one dataset are visually dissimilar compared to two randomly drawn images from another dataset: $\text{Div}_B(A) = 1 - p(d(a_1, a_2) < d(b_1, b_2))$ where $a_1, a_2 \in A$ and $b_1, b_2 \in B$ are drawn uniformly at random. Dataset $A$ is more diverse than $B$ if and only if $\text{Div}_B(A) > \text{Div}_A(B)$ . For $N$ datasets $A_1, A_2, \dots, A_N$ : $\text{Div}_{A_2,\dots,A_N}(A_1) = 1 - p\left(d(a_{11}, a_{12}) < \min_{i=2}^N d(a_{i1}, a_{i2})\right)$ where each pair $a_{i1}, a_{i2} \in A_i$ is drawn uniformly at random.
Knowl 3 — Empirical Density and Diversity Comparison Across Places, ImageNet, and SUN
empirical result
Relative density and diversity were evaluated across the 88 overlapping scene categories among the Places, ImageNet, and SUN datasets via Amazon Mechanical Turk (AMT) perceptual tests.

In the experiments:
- For the diversity evaluation, workers were presented with 12 image pairs (4 randomly drawn pairs from each database) per trial and asked to select the single most visually similar pair.
- For the density evaluation, each pair was constructed from a randomly chosen image and its 5th nearest neighbor in Gist descriptor Euclidean space (excluding the top 4 neighbors to prevent near-duplicates). Workers selected the most similar pair across 12 candidate pairs per trial.
Results across all 88 categories demonstrate:
1. Density: The average relative density is nearly indistinguishable across the three datasets.
2. Diversity: Relative diversity differs markedly, with Places exhibiting an average diversity score of $0.83$ , ImageNet scoring $0.67$ , and SUN scoring $0.50$ . In user rankings, pairs from SUN were selected as the most similar in $50\%$ of trials, whereas pairs from Places were judged most similar in only $17\%$ of trials. Categories showing the largest diversity spread include playground, veranda, and waiting room.

Knowl 4 — Places-CNN Classification Performance on Places 205 and SUN 205

data/table

Places-CNN uses the AlexNet architecture (the Caffe reference network) and is trained on 2,448,873 images across 205 categories (Places 205) using the Caffe framework for 300,000 iterations on an NVIDIA Tesla K40 GPU. The table below presents the direct Softmax output classification accuracy of Places-CNN against a linear SVM baseline trained on 4096-dimensional FC7 features from an ImageNet-trained CNN (ImageNet-CNN) on the test sets of Places 205 (41,000 images total; 200 per category) and SUN 205 (at least 50 images per category).

Places-CNN outperforms the ImageNet-CNN feature baseline by 9.2 percentage points on Places 205 and by 16.6 percentage points on SUN 205. The top-5 error rate of Places-CNN is $18.9\%$ on the Places 205 test set and $8.1\%$ on the SUN 205 test set.

Knowl 5 — Deep Feature Transfer Performance of Places-CNN vs ImageNet-CNN

data/table

The generic representation capabilities of scene-centric versus object-centric deep features were evaluated by extracting 4096-dimensional feature vectors from the Fully Connected Layer 7 (FC7) of Places-CNN (trained on Places 205) and ImageNet-CNN (trained on 1.2 million images of ILSVRC 2012 ImageNet). Both networks share the same AlexNet architecture. A linear SVM ( $C=1$ ) was trained on these features across four scene-centric benchmarks and four object-/action-/event-centric benchmarks.

Feature	SUN397	MIT Indoor67	Scene15	SUN Attribute
Places-CNN feature	54.32 $\pm$ 0.14%	68.24%	90.19 $\pm$ 0.34%	91.29%
ImageNet-CNN feature	42.61 $\pm$ 0.16%	56.79%	84.23 $\pm$ 0.37%	89.85%
Feature	Caltech101	Caltech256	Action40	Event8
Places-CNN feature	65.18 $\pm$ 0.88%	45.59 $\pm$ 0.31%	42.86 $\pm$ 0.25%	94.12 $\pm$ 0.99%
ImageNet-CNN feature	87.22 $\pm$ 0.92%	67.23 $\pm$ 0.27%	54.92 $\pm$ 0.33%	94.42 $\pm$ 0.76%

Places-CNN features outperform ImageNet-CNN features on all scene-centric datasets, establishing new state-of-the-art results on SUN397 ( $54.32\%$ ) and MIT Indoor67 ( $68.24\%$ ). Conversely, ImageNet-CNN features outperform Places-CNN features on object-centric (Caltech101, Caltech256) and action-centric (Stanford Action40) datasets, while both achieve comparable accuracy on UIUC Event8.

Knowl 6 — Hybrid-CNN Feature Transfer Across Scene and Object Benchmarks

data/table

Hybrid-CNN is trained on a combined dataset of 3.5 million images across 1,183 categories, formed by merging the training set of Places-CNN (Places 205) with the ImageNet training set after removing overlapping scene categories. Hybrid-CNN uses the same 8-layer architecture as Places-CNN and is trained for 700,000 iterations, achieving a validation accuracy of $52.3\%$ .

Classification accuracy/precision obtained using a linear SVM ( $C=1$ ) trained on 4096-dimensional FC7 features of Hybrid-CNN is summarized below:

Dataset	SUN397	MIT Indoor67	Scene15	SUN Attribute
Performance	53.86 $\pm$ 0.21%	70.80%	91.59 $\pm$ 0.48%	91.56%
Dataset	Caltech101	Caltech256	Action40	Event8
Performance	84.79 $\pm$ 0.66%	65.06 $\pm$ 0.25%	55.28 $\pm$ 0.64%	94.22 $\pm$ 0.78%

Combining scene and object training data enables Hybrid-CNN features to surpass both Places-CNN and ImageNet-CNN features on MIT Indoor67 ( $70.80\%$ ), Scene15 ( $91.59\%$ ), SUN Attribute ( $91.56\%$ ), and Stanford Action40 ( $55.28\%$ ).

Knowl 7 — Receptive Field Visualization of Scene-Centric vs Object-Centric CNN Units
empirical result
Internal unit representations across layers (Conv 1, Pool 2, Pool 5, and FC 7) were visualized and compared between ImageNet-CNN (object-centric) and Places-CNN (scene-centric) using a mean-activation receptive field procedure:
1. A combined set of 208,754 test images (100,000 from ImageNet LSVRC 2012 test set and 108,754 from SUN397) is provided as input to both networks.
2. For each unit at a given layer, the images are ranked by the magnitude of that unit's activation response.
3. The top 100 highest-activating images per unit are averaged to construct an empirical receptive field (RF) visualization.
4. Visualized unit averages are sorted by their first principal component for comparison.
Observations from the visualizations show:
- At Conv 1 (96 filters), both networks learn essentially identical low-level filters corresponding to oriented edges and color-opponent patches.
- Starting at Pool 2 ( $13 \times 13 \times 256$ ) and progressing through Pool 5 ( $6 \times 6 \times 256$ ) to FC 7 ( $4096 \times 1$ ), unit specializations diverge significantly. Units in ImageNet-CNN develop receptive fields tuned to localized object-like blobs, whereas units in Places-CNN develop receptive fields tuned to spatial structures, horizons, and landscape layouts.
Knowl 8 — Fine-Tuned Places-CNN Classification on SUN397
empirical result
Fine-tuning the entire Places-CNN on the SUN397 database using standard train/test splits and directly using the final softmax prediction layer yields a classification accuracy of $56.2\%$ .

This improves upon an ImageNet-CNN fine-tuned under the identical experimental protocol, which reaches $52.2\%$ , and substantially outperforms hand-crafted feature baselines on SUN397, such as combined kernel classifiers ( $37.5\%$ ) and Fisher Vector methods ( $47.20\%$ ).
Knowl 9 — Cross-Dataset Generalization and Bias Across Scene Subsets
empirical result
Cross-dataset generalization was evaluated across the 88 shared scene categories among SUN 88, ImageNet 88, and Places 88 using linear SVMs trained on ImageNet-CNN deep features.

Key empirical findings include:
1. Dataset Bias Effect: For any fixed training sample size per category ( $10^0$ to $10^2$ samples), training and testing within the same dataset yields higher accuracy than cross-dataset training and testing.
2. Large-Scale Data Effect: When all available training data is utilized:
  - Tested on SUN 88: Training on Places 88 achieves the highest accuracy ( $69.5\%$ ), exceeding training on SUN 88 ( $63.3\%$ ) and ImageNet 88 ( $62.8\%$ ).
  - Tested on Places 88: Training on Places 88 achieves $54.2\%$ , outperforming ImageNet 88 ( $44.6\%$ ) and SUN 88 ( $37.0\%$ ).
  - Tested on ImageNet 88: Training on ImageNet 88 achieves $65.6\%$ , outperforming Places 88 ( $60.3\%$ ) and SUN 88 ( $49.2\%$ ). Due to its scale and high visual diversity, training on Places 88 compensates for dataset bias and delivers superior cross-dataset transfer when evaluating on SUN 88.

Coverage note — No substantial contributed material was omitted.

References

1.P. Agrawal, R. Girshick, and J. Malik. Analyzing the performance of multilayer neural networks for object recognition. In Proc. ECCV. 2014.
2.Y. Bengio. Learning deep architectures for ai. Foundations and trends® in Machine Learning, 2009.
3.J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. CVPR, 2009.
4.C. Doersch, A. Gupta, and A. A. Efros. Mid-level visual element discovery as discriminative mode seeking. In In Advances in Neural Information Processing Systems, 2013.
5.J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. 2014.
6.R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. LIBLINEAR: A library for large linear classification. 2008.
7.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. Computer Vision and Image Understanding, 2007.
8.G. Griffin, A. Holub, and P. Perona. Caltech-256 object category dataset. 2007.
9.C. Heip, P. Herman, and K. Soetaert. Indices of diversity and evenness. Oceanis, 1998.
10.Y. Jia. Caffe: An open source convolutional architecture for fast feature embedding. http://caffe.berkeleyvision.org/, 2013.
11.T. Konkle, T. F. Brady, G. A. Alvarez, and A. Oliva. Scene memory is more detailed than you think: the role of categories in visual long-term memory. Psych Science, 2010.
12.A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In In Advances in Neural Information Processing Systems, 2012.
13.S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In Proc. CVPR, 2006.
14.Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation, 1989.
15.L.-J. Li and L. Fei-Fei. What, where and who? classifying events by scene and object recognition. In Proc. ICCV, 2007.
16.A. Oliva. Scene perception (chapter 51). The New Visual Neurosciences, 2013.
17.A. Oliva and A. Torralba. Modeling the shape of the scene: A holistic representation of the spatial envelope. Int'l Journal of Computer Vision, 2001.
18.G. Patterson and J. Hays. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In Proc. CVPR, 2012.
19.A. Quattoni and A. Torralba. Recognizing indoor scenes. In Proc. CVPR, 2009.
20.A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. arXiv preprint arXiv:1403.6382, 2014.
21.J. Sánchez, F. Perronnin, T. Mensink, and J. Verbeek. Image classification with the fisher vector: Theory and practice. Int'l Journal of Computer Vision, 2013.
22.E. H. Simpson. Measurement of diversity. Nature, 1949.
23.A. Torralba and A. A. Efros. Unbiased look at dataset bias. In Proc. CVPR, 2011.
24.J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Proc. CVPR, 2010.
25.B. Yao, X. Jiang, A. Khosla, A. L. Lin, L. Guibas, and L. Fei-Fei. Human action recognition by learning bases of action attributes and parts. In Proc. ICCV, 2011.

Citation

MLA

Zhou, B., et al. “Learning Deep Features for Scene Recognition Using Places Database”. Advances in Neural Information Processing Systems, vol. 27, 2014, pp. 487–95, https://proceedings.neurips.cc/paper_files/paper/2014/file/19ea3982b415d7bb3363917eb3d60c4a-Paper.pdf.

APA

Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., & Oliva, A. (2014). Learning Deep Features for Scene Recognition using Places Database. Advances in Neural Information Processing Systems, 27, 487–495. https://proceedings.neurips.cc/paper_files/paper/2014/file/19ea3982b415d7bb3363917eb3d60c4a-Paper.pdf

Chicago

Zhou, B., A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. 2014. “Learning Deep Features for Scene Recognition Using Places Database”. Advances in Neural Information Processing Systems 27: 487–95. https://proceedings.neurips.cc/paper_files/paper/2014/file/19ea3982b415d7bb3363917eb3d60c4a-Paper.pdf.

Harvard

Zhou, B. et al. (2014) “Learning Deep Features for Scene Recognition using Places Database”, Advances in Neural Information Processing Systems. Curran Associates, Inc., pp. 487–495. Available at: https://proceedings.neurips.cc/paper_files/paper/2014/file/19ea3982b415d7bb3363917eb3d60c4a-Paper.pdf.

Vancouver

1. Zhou B, Lapedriza A, Xiao J, Torralba A, Oliva A (2014) Learning Deep Features for Scene Recognition using Places Database. In: Advances in Neural Information Processing Systems. Curran Associates, Inc., pp 487–495

BibTeX

@inproceedings{NIPS2014_19ea3982,
 author = {Zhou, Bolei and Lapedriza, Agata and Xiao, Jianxiong and Torralba, Antonio and Oliva, Aude},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {Z. Ghahramani and M. Welling and C. Cortes and N. Lawrence and K. Weinberger},
 pages = {},
 publisher = {Curran Associates, Inc.},
 title = {Learning Deep Features for Scene Recognition using Places Database},
 url = {https://proceedings.neurips.cc/paper_files/paper/2014/file/19ea3982b415d7bb3363917eb3d60c4a-Paper.pdf},
 volume = {27},
 year = {2014}
}

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: Authors

Model	Places 205	SUN 205
Places-CNN (Softmax)	50.0%	66.2%
ImageNet-CNN feature + linear SVM	40.8%	49.6%

Learning Deep Features for Scene Recognition using Places Database

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Places Database Construction and Subsets

Knowl 2 — Formulations for Relative Dataset Density and Diversity

Knowl 3 — Empirical Density and Diversity Comparison Across Places, ImageNet, and SUN

Knowl 4 — Places-CNN Classification Performance on Places 205 and SUN 205

Knowl 5 — Deep Feature Transfer Performance of Places-CNN vs ImageNet-CNN

Knowl 6 — Hybrid-CNN Feature Transfer Across Scene and Object Benchmarks

Knowl 7 — Receptive Field Visualization of Scene-Centric vs Object-Centric CNN Units

Knowl 8 — Fine-Tuned Places-CNN Classification on SUN397

Knowl 9 — Cross-Dataset Generalization and Bias Across Scene Subsets

References

Citation

BibTeX

Source Code

Access the Paper