MobileNet-Yolo | MobileNetV2-YoloV3-Nano: 0.5BFlops 3MB HUAWEI P40: 6ms/img, YoloFace-500k:0.1Bflops 420KB:fire::fire | Computer Vision library

There are some questions when I read ssd-caffe code and I really need your help.

1. Native caffe only supports classification, data reading layer is commonly used to read LMDB database and read image for training

2. In order to support the input of multiple labels and input annotation boxes, I decide to use ssd-caffe, which adds an AnnotatedDataLayer layer to the native caffe. This newly added layer can support multiple labels and annotation boxes, but it has limitations. The reason is that the type of data it reads is still lmdb;

3. We now need to read the data of the data set randomly, but according to the query results, lmdb is a B+ tree structure, which can only be read sequentially through the iterator, so we want to change lmdb to read the images directly. However, the direct reading pictures of native caffe do not support multi-labels and annotation boxes. How can I modify the image_data_layers of caffe to support the input of annotation boxes (Can I follow AnnotatedDataLayer's approach to solve the problem)?

Note:

• Modified ssd-caffe source code: https://github.com/eric612/MobileNet-YOLO

• The file path of the newly added annotation box: /MobileNet-YOLO/src/caffe/layers/annotated_data_layer.cpp

• Native caffe file path for reading pictures directly: /MobileNet-YOLO/src/caffe/layers/image_data_layer.cpp