search

start-jsk / jsk_apc

Other
36 stars 35 forks source link

Current in-hand object recognition accuracy is low? BoF -> VGG16 #1486

Closed wkentaro closed 8 years ago

wkentaro commented 8 years ago

I tested with apc_object_recognition_bof.pkl.gz and apc_object_recognition_lgr.pkl.gz with 20% data of

The result was 52% (0.5244) with 25 objects.

Test program is here: https://github.com/wkentaro/vgg16/blob/master/examples/apc2015/test_bof.py

wkentaro commented 8 years ago

With VGG16 net 91% with 496k iteration, it exceeds 95%.

wkentaro commented 8 years ago

the label is estimated one and red is wrong.

wkentaro commented 8 years ago

Will be fixed via https://github.com/jsk-ros-pkg/jsk_recognition/pull/1688

wkentaro commented 8 years ago

Updated recognition result: 100 %

88%

wkentaro commented 8 years ago

On Gakka PC (T440p) CPU mode:

Input: /apply_mask_left_hand/output: 2.3hz output: /vgg16_object_recognitioin/output: 1.1hz

yuyu2172 commented 8 years ago

How slow is it on sheeta?

wkentaro commented 8 years ago

0.0000001 hz

yuyu2172 commented 8 years ago

I wonder why

wkentaro commented 8 years ago

I wonder why, too.

wkentaro commented 8 years ago

On doura, it is 0.5 hz faster than sheeta.

             topic                 rate    min_delta   max_delta   std_dev 
==========================================================================
/apply_mask_left_hand/output       2.756   0.237       0.903       0.189   
/vgg16_object_recognition/output   0.543   1.838       1.842       0.00
yuyu2172 commented 8 years ago

Oh good.