search

chuanqi305 / MobileNet-SSD

Caffe implementation of Google MobileNet SSD detection network, with pretrained weights on VOC0712 and mAP=0.727.
MIT License
2.03k stars 1.18k forks source link

Unable to detect objects relatively smaller in shape #51

Open linux-devil opened 6 years ago

linux-devil commented 6 years ago

I am trying to use above model , however when I detect smaller images it fails . When same object is passed into the network with larger size it is detected correctly. Any hint to make this model more robust to size of the objects?

lzx1413 commented 6 years ago
  1. the smallest step of mobilenet ssd is 16, so small object is hard to be detected.
  2. the last conv layers will generate feature maps with size 25611,so image which is smaller than 300 will not be processed with the whole network
linux-devil commented 6 years ago

https://github.com/balancap/SSD-Tensorflow This tensorflow model is able to detect smaller object as compared to caffe model.

lzx1413 commented 6 years ago

The smallest feature stride of a normal ssd model is 8

linux-devil commented 6 years ago

@lzx1413 thanks for the info

Where can I set the smallest step here?

ailihong commented 6 years ago

I have encountered the same problem. Have any good Suggestions and ideas? Thank you