1 / 8

DeeperVision and DeepInsight Solutions

This paper presents advancements in deep learning through the DeeperVision classification and DeepInsight detection frameworks. We utilize a Nesterov-based optimization technique to finely tune our models, achieving a top-1 accuracy improvement of 0.8%. The study explores data abstraction strategies, including stride and kernel size adjustments, and implements complex data augmentations and Spatial Pyramid Pooling (SPP). Our results demonstrate a top-5 error rate of 9.5% using an ensemble of five networks, outperforming traditional models such as GoogLeNet. We also seek postdoc and job opportunities.

ziv
Télécharger la présentation

DeeperVision and DeepInsight Solutions

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. DeeperVision and DeepInsight Solutions Junjie Yan*, Naiyan Wang*, Yinan Yu, Linjiao Zhao, Stan Z. Li, Dit-Yan Yeung * denotes equal contribution

  2. DeeperVision Classification • Deeper network always helps

  3. DeeperVision Classification • Nesterov method based optimization • With large momentum and Nesterov based optimization method, the algorithm could smooth out the optimization path. • It can improve top 1 accuracy by 0.8%

  4. DeeperVision Classification • More findings… • Slow down the speed of data abstraction (stride, kernel size, etc.) • More complicated data augmentations • Spatial Pyramid Pooling (SPP) • Our final results • Single net: Top 5 error: 10.5% • Ensemble 5 nets: Top 5 error: 9.5%

  5. Deep Insight Detection Region proposal + CNN feature extraction • Selective Search + Structural Edge [1] for region proposal. • 7/8/9 Convolution Layers + SPM +2 Fully Connected Layers. • Deeper Models need more tuning iterations. • Better (Deeper) Classification CNN always helps Detection. [1]C. Lawrence Zitnick and Piotr Dollár Edge Boxes: Locating Object Proposals from Edges ECCV 2014

  6. Diagnosis Experiments (on 2013-val2 ) Original RCNN 31.4 + 9conv + SPM 36.6 + more iterations 39.2 + Structural Edge Proposal 40.1 + 7/8/9 Conv Ensemble 40.7 + CLS Context 42.0

  7. Our Final Result • We have the best single model (40.2 mAP V.S. the 38.0 mAP of GoogLeNet) • We use a non-optimal ensemble method when submitting result. A better ensemble method leads to a 42.0 mAPon val2 after the competition. • Keeps improving…

  8. Advertisement • Junjie and I are looking for postdoc and job positions  • Junjie Yan: http://www.cbsr.ia.ac.cn/users/jjyan/main.htm • Naiyan Wang: http://winsty.net

More Related