일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | ||||||
2 | 3 | 4 | 5 | 6 | 7 | 8 |
9 | 10 | 11 | 12 | 13 | 14 | 15 |
16 | 17 | 18 | 19 | 20 | 21 | 22 |
23 | 24 | 25 | 26 | 27 | 28 |
- point cloud
- OpenCV
- reconstruction
- Object Detection
- 자료구조
- 3D
- deep learning
- NLP
- Threshold
- re-identification
- 프로그래머스
- 파이썬
- transformer
- 스택
- numpy
- 큐
- 임계처리
- attention
- cv2
- center loss
- Knowledge Distillation
- 알고리즘
- 딥러닝
- flame
- Computer Vision
- Deeplearning
- Python
- level2
- 논문 구현
- Object Tracking
- Today
- Total
목록딥러닝/3D (4)
공돌이 공룡의 서재
![](http://i1.daumcdn.net/thumb/C150x150/?fname=https://blog.kakaocdn.net/dn/basnPc/btrzCK9rp0G/7g80KXSGHzRL0qDUgQz7OK/img.png)
Introduction 사용할 수 있는 3D Face data 가 많지 않다는 것이 문제다. 그래서 이전 연구들은 surrogate shape label 과 같은 고전적인 방법들로 맞춰진 3D shape 이나 합성 데이터를 사용하는 방식을 택했다. 이 방식의 문제는 Domain gap 이 남아있다는 점이랑 label 이 완벽하지 않다는 점이다. 이런 이슈를 해결하기 위해서 라벨 없이 unsupervised 또는 weakly supervised 방식을 사용하는 연구들이 많이 나왔다. 이 연구들의 핵심은 미분 가능한 이미지를 형성하는 것인데, 얼굴 이미지를 네트워크의 예측 결과를 이용해서 렌더링하기 때문에 가능하다. 또한 입력 이미지와 렌더링된 이미지의 차이를 supervision (weakly 인듯) 으로..
![](http://i1.daumcdn.net/thumb/C150x150/?fname=https://blog.kakaocdn.net/dn/1ARo4/btrzyXa9abK/9kOyOtXv5bmvPjHH9tpWlK/img.png)
3D Face Reconstruction 에서 최근에 등장하여 3DMM 모델과 함께 많이 쓰이고 있다. 사실 FLAME 을 이용해서 어떻게 Reconstruction 을 할지가 더 중요하긴한데, 그래도 한번 살펴보자. 💡 FLAME = Faces Learned with an Articulated Model and Expressions : use 33000 scans : combines linear shape space with an articulated jaw, neck, and eyeballs, pose-dependent corrective blendshapes, and additional global expression blendshapes : low-dimensional but more expres..
![](http://i1.daumcdn.net/thumb/C150x150/?fname=https://blog.kakaocdn.net/dn/qSxQ5/btrdqRLc6gb/UVoyUCk9hxQkdVGkyhJCKK/img.png)
PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation https://arxiv.org/abs/1612.00593 글을 쓰는 시점에서 인용수가 5천이 넘었다. input으로 point cloud를 직접적으로 다룬 딥러닝 모델이다. 딥러닝 모델에서 point cloud를 다루기 위해서 무엇이 필요한지를 잘 설명하고 있다. Introduction point cloud는 정해진 format이 없기 때문에, 전에 있던 대부분의 연구가 이를 3D voxel grid로 바꾸거나 image들의 집합으로 바꾸고 난 후에 딥러닝 모델에 넣었다. 그러나 이렇게 처리하는 것은 데이터의 natural invariance를 가릴 수 있는 quan..
![](http://i1.daumcdn.net/thumb/C150x150/?fname=https://blog.kakaocdn.net/dn/mSyrv/btrduVx63z9/SDBOpwGBCrjrV84kCQiB81/img.png)
Expressive Body Capture: 3D Hands, Face, and Body from a Single Image https://arxiv.org/abs/1904.05866?utm_source=feedburner&utm_medium=feed&utm_campaign=Feed%3A+arxiv%2FQSXk+%28ExcitingAds%21+cs+updates+on+arXiv.org%29 [1] Introduction 사람의 행동을 더 잘 이해하기 위해선, 사람의 major keypoint들을 estimation 하는 것으로는 충분하지 않고, 몸, 손, 얼굴의 3D surface 정보들까지 포함해야 할 필요가 있다. 그런데 적절한 3D model과 3D training data가 충분하지 않아서, 이렇..