FCOS: Fully Convolutional One-Stage Object Detection
动机
- anchor free
- proposal free
- avoids the complicated computation related to anchor boxes
- calculating overlapping during training
- avoid all hyper-parameters related to anchor boxes
- size & shape
- positive／ignored／negative
- leverage as many foreground samples as possible
论点
anchor-based detectors
- detection performance is sensitive to anchor settings
- encounter difficulties in cases with large shape variations
- hamper the generalization ability of detectors
- dense propose：the excessive number of negative samples aggravates the imbalance
- involve complicated computation：such as calculating the IoU with gt boxes
FCN-based detector
- predict a 4D vector plus a class category at each spatial location on a level of feature maps
- do not work well when applied to overlapped bounding boxes
with FPN this ambiguity can be largely eliminated
anchor-free detector
- yolov1：only the points near the center are used，low recall
- CornerNet：complicated post-processing to match the pairs of corners
- DenseBox：difficulty in handling overlapping bounding boxes
this methos
- use FPN to deal with ambiguity
- dense predict：use all points in a ground truth bounding box to predict the bounding box
- introduce “center-ness” branch to predict the deviation of a pixel to the center of its corresponding bounding box
- can be used as a RPN in two-stage detectors and can achieve significantly better performance
方法
ground truth boxes，$B_i=(x_0, y_0, x_1, y_1, c)$，corners + cls
anchor-free：each location (x,y)，map into abs input image (xs+[s/2], ys+[s/2])
positive sample：if a location falls into any ground-truth box
ambiguous sample：location falls into multiple gt boxes，choose the box with minimal area
regression target：l t r b distance，location to the four sides
cls branch
- C binary classifiers
- C-dims vector p
- focal loss
- $\frac{1}{N_{pos}} \sum_{x,y}L_{cls}(p_{x,y}, c_{x,y}^*)$
calculate on both positive/negative samples
box reg branch
- 4-dims vector t
- IOU loss
- $\frac{1}{N_{pos}} \sum_{x,y}1_{\{c_{x,y}^>0\}}L_{reg}(t_{x,y}, t_{x,y}^)$
- calculate on positive samples
inference
choose the location with p > 0.05 as positive samples
two possible issues
- large stride makes BPR low, which is actually not a problem in FCOS
overlaps gt boxes cause ambiguity, which can be greatly resolved with multi-level prediction
FPN
- P3, P4, P5：1x1 conv from C3, C4, C5, top-down connections
P6, P7: stride2 conv from P5, P6
limit the bbox regression for each level
- $m_i$：maximum distance for each level
- if a location’s gt bbox satifies：$max(l^,t^,r^,b^)>m_i$ or $max(l^,t^,r^,b^)<m_{i-1}$，it is set as a negative sample，not regress at current level
- objects with different sizes are assigned to different feature levels：largely alleviate一部分box overlapping问题
for other overlapping cases：simply choose the gt box with minimal area
sharing heads between different feature levels
to regress different size range：use $exp(s_ix)$
- trainable scalar $s_i$
- slightly improve
center-ness
low-quality predicted bounding boxes are produced by locations far away from the center of an object
predict the “center-ness” of a location
normalized distance
sqrt to slow down the decay
[0,1] use bce loss
when inference center-ness is mutiplied with the class score：can down-weight the scores of bounding boxes far from the center of an object, then filtered out by NMS
- an alternative of the center-ness：use of only the central portion of ground-truth bounding box as positive samples，实验证明两种方法结合效果最好
architecture
- two minor differences from the standard RetinaNet
- use Group Normalization in the newly added convolutional layers except for the last prediction layers
- use P5 instead of C5 to produce P6&P7
- two minor differences from the standard RetinaNet