Pre-trained models

There are several models that are pre-trained and made available. All these models are trained on the COCO dataset and can be used for detecting the objects that are available in the COCO dataset such as humans and cars. These models are also useful for transfer learning for a new task such as traffic sign detection. A table of pre-trained models is shown here with relative speed and mAP on the COCO dataset. Various algorithms are trained with different CNN and are depicted in the names:

Model name	Speed	COCO mAP
ssd_mobilenet_v1_coco	fast	21
ssd_inception_v2_coco	fast	24
rfcn_resnet101_coco	medium	30
faster_rcnn_resnet101_coco	medium	32
faster_rcnn_inception_resnet_v2_atrous_coco	slow	37

 

Based on the requirement, you can choose from the model. Download the SSD model trained on Mobilenet and extract it as shown here by going to the working directory:

mkdir Chapter04 && cd Chapter04
wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_11_06_2017.tar.gz
tar -xzvf ssd_mobilenet_v1_coco_11_06_2017.tar.gz

There will be various files in the Chapter04 folder, which are listed here:

• The is the proto-definition of the graph—graph.pbtxt 
• The weights of the graph frozen and can be used for inference—frozen_inference_graph.pb
• Checkpoint files
  • model.ckpt.data-00000-of-00001
  • model.ckpt.meta
  • model.ckpt.index

This model will be used in the next section for detection tasks.