2015:
single shot detector(SSD)
YOLO: you only look once
### Fast Yolo
- 9 of 24 CNN layers from GoogleNet
- fully connected detector network
- real-time systems on PASCAL VOC 2007
-
2016:
YOLO9000(v2) :Better, Faster, Stronger
- backbone: darknet-19(19 conv layers and 5 maxpooling layers)
2017:
Fast YOLO:
- optimzided yolo v2 reduced parameters,evolutionary deep intelligence framework
- introduce a motion-adaptive-inference method.(intuitively not very reliable)
2018:
YOLO v3:
- backbone: Darknet-53(53 conv layers)
- 1/2 fast as YOLO v2.
YOLO-lite
- just modify tiny yolo and see how it worked.
tiny ssd:
- Inspired by the **Fire microarchitecture ** from SqueezeNet. First subnet stack of Tiny SSD as a standard convolutional layer followed by a set of highly optimized Fire modules.
- determine the ideal number of Fire modules as well as the ideal microarchitecture.
- determined empirically that 10 Fire modules provide strong object detection performance
- feature layer:
- result
- half precision floating-point parameters.
2019:
EfficientDet
- speed: same level as yolo3
# 2020: ## YOLO v4:
#THings need checking out: small backbone: SqueezeNet, ShuffleNet Two_stage detection head:
- RCNN,fast-RCNN,faster-RCNN One stage detection:
- retinaNet Neck layers:
-
PAN,FPN,BiFPN
- what exactly is receptive field?
