Skip to content

sureone/darknet_faster_rcnn

 
 

Darknet Logo

Darknet

Darknet is an open source neural network framework written in C and CUDA. It is fast, easy to install, and supports CPU and GPU computation.

Faster_rcnn model added under Darknet

[fasterrcnn]
classes=20
rois_nms_thresh=0.7
rois_min_area_thresh=16
train_rcnn_flg=1
rpn_sample_num=256
rpn_sample_pos_ratio=0.25
rpn_iou_pos_thresh=0.7
rpn_iou_neg_thresh=0.3
downsample_ratio=16
anchor_scale=4, 8 ,16,32
anchor_ratio=0.5,1,2
train_pre_nms_num=12000
train_post_nms_num=2000
test_pre_nms_num=6000
test_post_nms_num=300
rois_sample_num=128
rois_sample_ratio=0.25
roialign_pooling_height=7
roialign_pooling_width=7
pos_iou_thresh=0.5
neg_iou_thresh_hi=0.5
neg_iou_thresh_lo=0

How to train and test faster_rcnn?

train on pretrained resnet50 model

./darknet faster_rcnn train cfg/voc.data cfg/resnet50_faster_rcnn.cfg resnet50.weights -pretrain

Using -pretrain, the model will initialize the rpn and rcnn randomly.

imageNet pretrained resnet50 and resnet152 model download link

https://pjreddie.com/media/files/resnet50.weights
https://pjreddie.com/media/files/resnet152.weights

continue training the model

./darknet faster_rcnn train cfg/voc.data cfg/resnet50_faster_rcnn.cfg backup/resnet50_faster_rcnn.backup

test the model

./darknet faster_rcnn test cfg/voc.data cfg/resnet50_faster_rcnn.cfg backup/resnet50_faster_rcnn.backup /to/your/image/path/.jpg

Future work

  1. adding the nms_gpu function for the rois nms
  2. FPN structure will be added soon
  3. adding focal loss function for unbalanced fg and bg
  4. disable the delta and weights_update memory calloc when the only inference using
  5. testing the performance on the VOC datasets

Project modified by Hao. March. 2020

Releases

No releases published

Packages

No packages published

Languages

  • C 88.5%
  • Cuda 9.6%
  • Other 1.9%