Hi Developer:
First of all, I need to confirm: "Does SNPE original layers support multi-objects detection by FasterRCNN?"
Here is my experience. I have tried mobilenet_ssd both in snpe SDK and 820a board succeed. And I can run vgg16_faster_rcnn_final.caffemodel by snpe SDK. But I always get one output (21 for cls_prob and 84 for bbox_pred) regardless of the number of “max_num_rois”, which is a parameter you can set before the converting operation from *.caffemodel to *.dlc file.
According to snpe’s Limitations and Issues: Only batch of 1 is supported for ROIPooling layer, So I made an experiment, I made a udl layer named “MyROIPool”, which can support batch > 1. After a lot of coding operation, I can get a multi-objects detection by using vgg16_faster_rcnn_final_MyROIPool.caffemodel, and the results is normal.
Hopefully, I have made myself understood!
thanks!
Hi zakai:
I also met the same problem. I am writing a UDL for ROIPooling layer. Only one question: How to get the real number of rois in my UDL?
Because the real number of rois varies during process images, but udl's setup function is running before processing images. I cannot get the real number of rois in execute function of ROIPooling.
Thansks in advance!