Community
cancel
Showing results for 
Search instead for 
Did you mean: 
idata
Community Manager
363 Views

Retraining with fewer classes

Hi

 

So I took the COCO SSD 300 model from here:

 

https://github.com/weiliu89/caffe/tree/ssd

 

File here:

 

https://drive.google.com/open?id=0BzKzrI_SkD1_dUY1Ml9GRTFpUWc

 

I changed the output from:

 

layer {

 

name: "detection_out"

 

type: "DetectionOutput"

 

bottom: "mbox_loc"

 

bottom: "mbox_conf_flatten"

 

bottom: "mbox_priorbox"

 

top: "detection_out"

 

include {

 

phase: TEST

 

}

 

detection_output_param {

 

num_classes: 81

 

share_location: true

 

background_label_id: 0

 

nms_param {

 

nms_threshold: 0.45

 

top_k: 400

 

}

 

save_output_param {

 

label_map_file: "data/coco/labelmap_coco.prototxt"

 

}

 

code_type: CENTER_SIZE

 

keep_top_k: 200

 

confidence_threshold: 0.01

 

}

 

}

 

To this:

 

layer {

 

name: "detection_out"

 

type: "DetectionOutput"

 

bottom: "mbox_loc"

 

bottom: "mbox_conf_flatten"

 

bottom: "mbox_priorbox"

 

top: "detection_out"

 

include {

 

phase: TEST

 

}

 

detection_output_param {

 

num_classes: 81

 

share_location: true

 

background_label_id: 0

 

nms_param {

 

nms_threshold: 0.45

 

top_k: 100

 

}

 

code_type: CENTER_SIZE

 

keep_top_k: 100

 

confidence_threshold: 0.3

 

}

 

}

 

The accuracy seems low but the main issue is it runs at about 1.5 seconds per frame. Any ideas why?

 

Is it because there are the full 91 classes?

 

To get the good performance of the 20 class SSD Demo file will I need to retrain with 20 classes only?

 

Cheers,

 

Dan

0 Kudos
0 Replies