CODEWITHSUNDEEP
tensorflowobject-detection
Xxcc
Xxcc

Reputation: 91

Don't need some existed classes in pre-trained models

Tendorflow object detection.

I know you can train new class with existed pre-trained models in here:

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

But I want to only detect 'person' and 'car',these two classes already existed,futher training is not neccessary.

However,model such as 'faster_rcnn_inception_v2_coco',It detect 90 classes,and runs too slowly for me.

Is there a way to cut down the classes,so the detection runs faster?

PS:I've filtered COCO dataset to one that only has 'person' and 'car',but following the common steps I googled,I got bad training result.I can only successfully detect 'person'.Wander what's the reason.

Upvotes: 5

Views: 2125

Answers (1)

Xxcc
Xxcc

Reputation: 91

Solved myself.If you use 'person' or 'car' in coco,which id is 1 and 3 in pre-trained model,and that cannot be changed.

In .config,set 'num_classes'=3,although your classes num is 2.

The 3 is to fit 'car'-id.

(e.g.If you choose 2 classes,'person'=1,'truck'=8,your 'num_classes' should be 8)

Also,choose your epoch properly,you can see:http://neuralnetworksanddeeplearning.com/chap3.html

Upvotes: 1

Related Questions