CODEWITHSUNDEEP
pythoncaffeobject-detectionface-detectionlmdb
yogeshmurthy_
yogeshmurthy_

Reputation: 41

Generating lmdb files for object detection in Caffe

I am trying to train an SSD-based face detector from scratch in Caffe. I have a dataset of images where the bounding boxes for faces are stored in a csv file. In all the tutorials and code I've come across so far, convert_annoset tool is used to generate an lmdb file for object detection. However, in the latest Caffe version on Github (https://github.com/BVLC/caffe), this tool has been removed.

The two options I see to deal with this issue are:

  1. Rewrite the convert_annoset tool using functions in the current Caffe library

  2. Use other python packages (such as lmdb and OpenCV) to manually create lmdb files from the images and bounding box information

I've been trying to rewrite the code but I am unable to find certain classes and functions that were used in the original code such as AnnotatedDatum_AnnotationType and LabelMap in the current version.

Any suggestions for how to proceed with creating lmdb for the object detection problem?

EDIT: I've just realized that AnnotatedData layer no longer exists in the master branch of Caffe. Does this mean that detection is not possible in this version? Do I have to use some older fork such as https://github.com/weiliu89/caffe for detection or is there any other option?

Upvotes: 4

Views: 769

Answers (1)

Hong-333
Hong-333

Reputation: 81

Since the original Caffe does not have a layer for object detection(mobilenet-ssd)

so use this Caffe (https://github.com/weiliu89/caffe) ssd branch

This Caffe has a convert_annoset that can generate an lmdb.

Upvotes: 0

Related Questions