Open junxnone opened 4 years ago
P3~P7 -> Prediction Head
Prototype mask * coefficient = final mask
4 + C + k
anchor + class + mask coefficient
c positive labels + 1 background label
neg:pos = 3:1
hxwxk
nxk
n
assembled masks M
GT masks
pixel-wisebinary cross entropy
c x n x n
c classes
top n detections
IoU
t
feature map
1x1 conv layer + sigmoid
c channel mask
cropped mask
ROI Align layer
FC Layer
+1.2 ms
-1.4 fps
原版 MS RCNN MaskIoU Head
+28 ms
-16.9 fps
Throughput
Deformable Conv Module
C3 ~ C5
3x3 Conv Layer
3x3 Deformable Conv Layer
+1.8 mAP
+8 ms
interval = 3
+2.8 ms
+1.6 mAP
+5/3x anchors
+3x anchors
@COCO test-dev
@Titan Xp
550x550
YOLACT/YOLACT++
Arch
Protonet
Head
4 + C + k
=anchor + class + mask coefficient
c positive labels + 1 background label
neg:pos = 3:1
Mask Assembly
hxwxk
prototyep masksnxk
mask coefficients - NMS 和 Threshold 后剩n
instanceLoss
assembled masks M
> masks
pixel-wisebinary cross entropy
Fast NMS
c x n x n
IoU 矩阵c classes
: c 个分类top n detections
: 按 score 排序后的 n 个 BBoxIoU
大于t
的 BBoxSemantic Segmentation Loss - extra losses
feature map
+1x1 conv layer + sigmoid
=c channel mask
Crop & Threshold
YOLACT++
Fast Mask Re-Scoring Network
cropped mask
ROI Align layer
FC Layer
+1.2 ms
Throughput-1.4 fps
原版 MS RCNN MaskIoU Head
, 则+28 ms
-16.9 fps
Deformable Convolution with Intervals
Throughput
, 不使用整个Deformable Conv Module
, 只替换C3 ~ C5
中的3x3 Conv Layer
为3x3 Deformable Conv Layer
-+1.8 mAP
-+8 ms
interval = 3
-+2.8 ms
+1.6 mAP
Optimized Prediction Head
+5/3x anchors
+3x anchors
YOLACT++ Improvements
YOLACT/YOLACT++ Performance vs Others Network
@COCO test-dev
@Titan Xp
550x550
存在的问题
Reference