1
Reference A. Khosla, N. Jayadevaprakash, B. Yao and L. FeiFei. “Novel Dataset for FineGrained Image Categoriza:on.” First Workshop on FineGrained Visual Categoriza8on, CVPR 2011. Experiments – Stanford Dogs Dataset Example Images Stanford University 15 vs 100 Training Examples/Class Accuracy wrt Number of Training Examples 13.8% 17.7% 21.2% 24.1% 26.1% 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Different Age Different pose ParHal Occlusion Different classes, very similar appearance Reference image: Chihuahua Chihuahua Chihuahua Chihuahua Whippet Toy Terrier West Highland Terrier Different Appearance Chihuahua Toy Terrier Mo:va:on Our Work Collected/cleaned and annotated a novel dataset addressing the drawbacks of exisHng datasets Provide baseline results on the datasets that demonstrate their effecHveness The Vision Lab Computer Science Dept. MoHvaHon: The hierarchical structure and image distribuHon of ImageNet make it a great resource for finegrained image categorizaHon datasets Our contribu:ons: Details of Proposed Datasets LabelMe 85 classes of object: >500 im/class 211 classes of object: >100 im/class 6570 classes of object: >500 im/class 9836 classes of object: >100 im/class New Datasets No. of Classes No. of Images Images per class Training Images/class Visibility varies? Bounding boxes? Stanford Dogs 120 20580 ~180 100 Yes Yes Previous Datasets No. of Classes No. of Images Images per class Training Images/class Visibility varies? Bounding boxes? CUB200 200 6033 30 15 Yes Yes PPMI 24 4800 200 100 No Yes PASCAL AcHon 9 1221 135 ~60 Yes Yes Dataset Collec:on Our dataset significantly increases the number of images to allow algorithms to beaer capture the subtle differences between finegrained classes bounding boxes provided for all images Conclusions Finegrained image categorizaHon can greatly benefit from a large number of training examples The proposed datasets can adequately address the drawbacks in exisHng datasets 15 Images/class 100 Images/class Novel Dataset for FineGrained Image Categoriza:on Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao and Li FeiFei {aditya86,bangpeng,feifeili}@cs.stanford.edu, [email protected] Note: Above results are on 70 of 120 classes as collecHon is in progress (check website for latest results)

Novel’Dataset’for’Fine2Grained’Image’Categoriza:on’vision.stanford.edu/documents/KhoslaJayadevaprakashYaoFeiFei_F… · 21.2% 24.1% 26.1% 00.10.20.30.40.50.60.7 Differentpose#

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Novel’Dataset’for’Fine2Grained’Image’Categoriza:on’vision.stanford.edu/documents/KhoslaJayadevaprakashYaoFeiFei_F… · 21.2% 24.1% 26.1% 00.10.20.30.40.50.60.7 Differentpose#

Reference  A.  Khosla,  N.  Jayadevaprakash,  B.  Yao  and    L.  Fei-­‐Fei.  “Novel  Dataset  for  Fine-­‐Grained  Image  Categoriza:on.”    First  Workshop  on  Fine-­‐Grained  Visual  Categoriza8on,  CVPR  2011.  

Experiments  –  Stanford  Dogs  Dataset  Example  Images  

Stanford  University  

15    vs  100  Training  Examples/Class  

Accuracy  wrt  Number  of  Training  Examples  

13.8%

17.7%

21.2%

24.1%

26.1%

0                0.1            0.2            0.3          0.4            0.5            0.6              0.7  

Different  Age  Different  pose   ParHal  Occlusion  

Different  classes,    very  similar  appearance  

Reference  image:  Chihuahua  

Chihuahua  Chihuahua   Chihuahua  

Whippet   Toy  Terrier  West  

Highland  Terrier  

Different  Appearance  

Chihuahua  

Toy  Terrier  

Mo:va:on  

Our  Work  

§   Collected/cleaned    and  annotated  a  novel  dataset  addressing  the  drawbacks  of  exisHng  datasets  §   Provide  baseline  results  on  the  datasets  that  demonstrate  their  effecHveness  

The  Vision  Lab  Computer  Science  Dept.  

•   MoHvaHon:  The  hierarchical  structure  and  image  distribuHon  of  ImageNet  make  it  a  great  resource  for  fine-­‐grained  image  categorizaHon  datasets  •   Our  contribu:ons:  

Details  of  Proposed  Datasets  

LabelMe  

85  classes  of  object:  >500  im/class  211  classes  of  object:  >100  im/class  

6570  classes  of  object:  >500  im/class  9836  classes  of  object:  >100  im/class  

New  Datasets  

No.  of  Classes  

No.  of  Images  

Images  per  class  

Training  Images/class  

Visibility  varies?  

Bounding  boxes?  

Stanford  Dogs   120   20580   ~180   100   Yes   Yes  

Previous  Datasets  

No.  of  Classes  

No.  of  Images  

Images  per  class  

Training  Images/class  

Visibility  varies?  

Bounding  boxes?  

CUB-­‐200   200   6033   30   15   Yes   Yes  PPMI   24   4800   200   100   No   Yes  PASCAL  AcHon   9   1221   135   ~60   Yes   Yes  

Dataset  Collec:on  

•     Our  dataset  significantly  increases  the  number  of  images  to  allow  algorithms  to  beaer  capture  the  subtle  differences  between    fine-­‐grained  classes  

bounding  boxes  provided  for  all  images  

Conclusions  

•   Fine-­‐grained    image  categorizaHon  can  greatly  benefit  from    a  large  number  of  training  examples  •   The  proposed  datasets  can  adequately  address  the  drawbacks  in  exisHng  datasets  

15  Images/class  100  Images/class  

Novel  Dataset  for  Fine-­‐Grained  Image  Categoriza:on  Aditya  Khosla,  Nityananda  Jayadevaprakash,  Bangpeng  Yao  and    Li  Fei-­‐Fei  

{aditya86,bangpeng,feifeili}@cs.stanford.edu, [email protected]

Note:  Above  results  are  on  70  of  120  classes  as  collecHon    is  in  progress  (check  website  for  latest  results)