MearcAll papers

arXiv 1409.1556 · cs.CV

Very Deep Convolutional Networks for Large-Scale Image Recognition

Karen Simonyan, Andrew Zisserman

Claims were extracted automatically by reducer-1 on 2026-11-12. The paper's license does not allow republishing it, so long claims are shortened to a quotation. Read the full text on arXiv.

No reproduction recordedres_236471ab
Our main contribution is a thorough evaluation of networks of increasing depth using an architecture with very small (3 × 3) convolution filters, which shows that a significant improvement on … (excerpt)

ABSTRACT · page 1 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_8b5bdf00
As a result, we come up with significantly more accurate ConvN et architectures, which not only achieve the state-of-the-art accuracy on ILSVRC classifica tion and localisation tasks, but are also … (excerpt)

I NTRODUCTION · page 1 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_c6455f00
4, such normalisation does not improve the performance on the ILSVRC dataset, but l eads to increased memory consumption and computation time.

A RCHITECTURE · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_2d9549e5
The width of conv. layers (the number of channels) is rather small, sta rting from 64 in the first layer and then increasing by a factor of 2 after each … (excerpt)

C ONFIGURATIONS · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_b62c9b4b
The learning rate was initially set to 10−2, and then decreased by a factor of 10 when the validation set accuracy stopped improving.

T RAINING · page 4 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_df145a24
4, using several values of Q for each S leads to improved performance).

T ESTING · page 5 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_12073591
Notably, in spite of the same de pth, the configuration C (which contains three 1 × 1 conv. layers), performs worse than the configuration D, which uses 3 × 3 … (excerpt)

S INGLE SCALE EVALUATION · page 6 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_39ef9449
The error rate of our arc hitecture saturates when the depth reaches 19 layers, but even deeper models might be beneficial for largerdatasets.

S INGLE SCALE EVALUATION · page 6 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_8fffbfab
The top-1 error of the shallow net wa s measured to be 7% higher than that of B (on a center crop), which confirms that a deep net with smal … (excerpt)

S INGLE SCALE EVALUATION · page 6 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_d18cbd9c
Finally, scale jittering at training time ( S ∈ [256; 512]) leads to significantly better results than training on images with fixed smallest side ( S = 256 or S … (excerpt)

S INGLE SCALE EVALUATION · page 6 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_d6897fbe
6 Published as a conference paper at ICLR 2015 The results, presented in Table 4, indicate that scale jittering at test time leads to better performance (as compared to evaluating … (excerpt)

M ULTI -SCALE EVALUATION · page 6 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_ce9322d7
On the test set, the configuration E achieves 7.3% top-5 error.

M ULTI -SCALE EVALUATION · page 7 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_9f97c7d2
As can be seen, using multiple crops performs sl ightly better than dense evaluation, and the two approaches are indeed complementary, as their combination outperforms each of them.

M ULTI -CROP EVALUATION · page 7 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_9fc7ff3c
For refere nce, our best-performing single model achieves 7.1% error (model E, Table 5).

C ONV NET FUSION · page 7 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_2ed33d97
As can be seen from Table 7, our very deep ConvNets significantly outperform the previous generation of models, which achieved the best results in the ILSVR C-2012 and ILSVRC-2013 competitions.

C OMPARISON WITH THE STATE OF THE ART · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_e6d4b0c1
Our result is also competitive with respect to the cla ssification task winner (GoogLeNet with 6.7% error) and substantially outperforms the ILSVRC-2013 winn ing submission Clarifai, which achieved 11.2% with … (excerpt)

C OMPARISON WITH THE STATE OF THE ART · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_0a89ad14
In terms of the single-net performance , our architecture achieves the best result (7.0% test error), outperforming a single GoogLeNet by 0.9%.

C OMPARISON WITH THE STATE OF THE ART · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_47b76fdc
In the appendix, we also show that our models generalise well to a wide range of tasks and datasets, matching or outperformin g more complex recognition pipelines built around less … (excerpt)

C ONCLUSION · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_2e9fbcfe
As can be seen from Table 9, application of the localisation ConvNet to the whole image substantially improves the results compared to using a center crop (Table 8), despite using … (excerpt)

L OCALISATION EXPERIMENTS · page 11 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_154b9d93
4), testing at several scales and combining the predictions of multiple networks further improves the performance.

L OCALISATION EXPERIMENTS · page 11 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_7abeda1e
Following that line of work, we investigate if our models lead to better performance than more shallow models utilised in the state-of-the-art methods.

G ENERALISATION OF VERY DEEP FEATURES · page 12 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_457d53cb
It is worth noting though that the improvement over a smalle r range of {256, 384, 512} was rather marginal (0.3%).

G ENERALISATION OF VERY DEEP FEATURES · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_b6754b42
Our methods set the new state of the art across image representations, pretrained on the ILSVRC dataset, outperforming the previous b est result of Chatfield et al. (2014) by more … (excerpt)

G ENERALISATION OF VERY DEEP FEATURES · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_8f8e1478
As can be seen, the deeper 19-layer Net-E performs better than the 16-layer Net-D, and their combination further improves the performance.

G ENERALISATION OF VERY DEEP FEATURES · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_49cbecee
On Caltech-256, our features outperform the state of the art (Chatfield et al., 2014) by a large margin (8.6%).

G ENERALISATION OF VERY DEEP FEATURES · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_b4c2a409
Since the public release of our models, they have been active ly used by the research community for a wide range of image recogniti on tasks, consistently outperforming more shallow … (excerpt)

G ENERALISATION OF VERY DEEP FEATURES · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_71c947fc
(2014) achieve the state of the object detection results by replacing the ConvNet of Krizhe vsky et al. (2012) with our 16-layer model.

G ENERALISATION OF VERY DEEP FEATURES · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

Reproduced one of these claims?

Record your result in the Mearc app on your own machine, then use Publish this result with this paper's id (1409.1556) and the claim id shown next to the claim. Mearc creates a file that holds only what you choose to publish. Add it to the public claim data by pull request; a maintainer other than you reviews it before it counts.