densepose spotlight plumsix - cocodataset.org
TRANSCRIPT
![Page 1: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/1.jpg)
Deep DensePose R-CNNPlumSix
x
![Page 2: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/2.jpg)
About PlumSix
Byeonguk Min, KAIST [email protected]
Hakyeong Kim, KAIST [email protected]
Jaehwee Lee, KAIST [email protected]
![Page 3: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/3.jpg)
Mentors
SHOUNAN [email protected]
Seungje [email protected]
Youngbak [email protected]
of Game Dev. AI Team, nARC(netmarble AI Revolution Center), netmarble
![Page 4: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/4.jpg)
Task
![Page 5: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/5.jpg)
Origin DensePose R-CNN
• We focused on that• Output resolution isn’t large enough• Time complexity doesn’t matter in the evaluation→ Approach : build up-sampling layers deeper
FPNDense Regression (Segmentation + UV mapping)
…
Feature Vector(*, 256, 14, 14)
Convolution stage[3x3 Conv + Relu]x8
Deconvolution stage
AnnIndex
Index_UV
U_estimated
V_estimated
![Page 6: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/6.jpg)
Our Model
![Page 7: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/7.jpg)
Inspiration – FCN8
![Page 8: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/8.jpg)
Our Model
![Page 9: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/9.jpg)
Experiments
• Fine-tuned DensePose R-CNN (+X101-32x8d)• Most of hyper-parameters followed baseline’s
• Image per minibatch : 3 → 2• Learning rate x0.666• Learning schedule x1.5 (195k iter)• Used Xavier initializer for new layers
• No ensembles• No additional datasets• Freeze backbone, faster branch
![Page 10: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/10.jpg)
Results
![Page 11: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/11.jpg)
Conclusion
• Our model is nothing but fine-tuned deep DensePose R-CNN which returns higher resolution output• We feed FPN layers again• mAP performs about 2% better• But for smaller area, our model doesn’t help
• We may try some techniques introduced in the DensePosepaper
![Page 12: densepose spotlight plumsix - cocodataset.org](https://reader031.vdocuments.mx/reader031/viewer/2022013012/61cefd54217d765910589d3e/html5/thumbnails/12.jpg)
Thank you