【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints
0 likes647 views
The document discusses advanced methodologies in object detection, particularly focusing on the effectiveness of corner pooling techniques and heatmap embeddings in neural networks. It references several academic papers and results from the COCO dataset evaluations. Key metrics like average precision (AP) at different IoU thresholds are presented to compare one-stage and two-stage detection frameworks.
![be r
• -
– 21.YiT n I N aLs
– - C A 2 00 R A 9 Y g
– be 2 006 - C 2 006
- C A 2 006 2 - 0 6 C
• -
– ][ 21.h Fs Y
– D C oSO FtcRw
– be 51 16 33,6 51 1 6
,33, 6 2 C 0 C 6 C
-
D A](https://image.slidesharecdn.com/daikikimuraeccv20182-181014142755/85/ECCV-2018-CornerNet-Detecting-Objects-as-Paired-Keypoints-3-320.jpg)
![w o T
• ,,3 ( 4 )
– A 7 ap V
0 an[
2D C 9 21 bh
– R a, A 9A
-9C9 C AV] Sn[
O alVg
21 r ai
– bhVt S
Vcx bh
I 21 bh
o , A 9A aesVo
[arxiv: 1703.10295]](https://image.slidesharecdn.com/daikikimuraeccv20182-181014142755/85/ECCV-2018-CornerNet-Detecting-Objects-as-Paired-Keypoints-4-320.jpg)
![nr TcE S M T
• 0 1
– fTn mT l gNT P l o
– [ [T
n
55
A
vMO C
[ T S C]dEbe] p
– , , T aSO , pt
g C Is S Ol
SC 7 2 i
[arxiv: 1611.05424]](https://image.slidesharecdn.com/daikikimuraeccv20182-181014142755/85/ECCV-2018-CornerNet-Detecting-Objects-as-Paired-Keypoints-5-320.jpg)
![Hourglass network [28, ECCV2016, arxiv: 1603.06937]
[28]
hourglass
SOTA
CournerNet](https://image.slidesharecdn.com/daikikimuraeccv20182-181014142755/85/ECCV-2018-CornerNet-Detecting-Objects-as-Paired-Keypoints-7-320.jpg)
![N P
• -
• . C , 8 . C 2 3
• a
– /G0 b H ] L[
– 0:C 1 X Tc
:C 5 CC 8:C 5 : C . C
3.5
! = # = 0.1 ' = 1](https://image.slidesharecdn.com/daikikimuraeccv20182-181014142755/85/ECCV-2018-CornerNet-Detecting-Objects-as-Paired-Keypoints-13-320.jpg)
【ECCV 2018】CornerNet: Detecting Objects as Paired Keypoints
- 1. : . K , ):C C J: HM D / C R N U O / : : P HH D H I H D C:H D : H C :L J : AA:C : HBA (D C: :H ):H: H C 1 : H : :M D CH
- 2. • () ( A • 2 • 2 () ( A C - A • A
- 3. be r • - – 21.YiT n I N aLs – - C A 2 00 R A 9 Y g – be 2 006 - C 2 006 - C A 2 006 2 - 0 6 C • - – ][ 21.h Fs Y – D C oSO FtcRw – be 51 16 33,6 51 1 6 ,33, 6 2 C 0 C 6 C - D A
- 4. w o T • ,,3 ( 4 ) – A 7 ap V 0 an[ 2D C 9 21 bh – R a, A 9A -9C9 C AV] Sn[ O alVg 21 r ai – bhVt S Vcx bh I 21 bh o , A 9A aesVo [arxiv: 1703.10295]
- 5. nr TcE S M T • 0 1 – fTn mT l gNT P l o – [ [T n 55 A vMO C [ T S C]dEbe] p – , , T aSO , pt g C Is S Ol SC 7 2 i [arxiv: 1611.05424]
- 7. Hourglass network [28, ECCV2016, arxiv: 1603.06937] [28] hourglass SOTA CournerNet
- 9. 13 2 1 1 !"#$: %, ' class ( )"#$: %, ' class ( *: +, ,: + = 2 , , = 4 Downsampling offset
- 10. (1, 02 7 ) 6 1, 02 !"#$$/!"#&' ()* /(+* E , 7 1, 02 (- ()* :(+* L ∆ ∆= 1
- 11. 1 / max pooling
- 12. 1 2
- 13. N P • - • . C , 8 . C 2 3 • a – /G0 b H ] L[ – 0:C 1 X Tc :C 5 CC 8:C 5 : C . C 3.5 ! = # = 0.1 ' = 1
- 14. AP50 AP70 http://cocodataset.org/#detection-eval AP One-stage SOTA Two-stage competitive
- 15. ⇒ Corner pooling
- 17. • ) (( , l L O Y D ! • t i D 1 - , O O 7 Op SD 7 an sg • Oe D o O