CS-F441: Selected Topics from Computer Science (Deep ... › dl › lecture-12-kt-cv.pdf2 Cite 6375...

CS-F441: SELECTED TOPICS FROM COMPUTER

SCIENCE (DEEP LEARNING FOR NLP & CV)

Lecture-KT-14: Segmentation (U-Net), Object Detection (Yolo)

Dr. Kamlesh Tiwari,Assistant Professor,

Department of Computer Science and Information Systems,BITS Pilani, Rajasthan-333031 INDIA

Nov 20, 2019 (Campus @ BITS-Pilani July-Dec 2019)

Segmentation

ISBI challenge for segmentation of neuronal structures in electronmicroscopic stacksWorks with very few training images (30/application) and touchingboundary. Yield more precise segmentationData augmentation is essential (mainly shift, rotation and elasticdeformation)

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 2 / 8

U-Net1

ISBI DIC-HeLa achieved 77.6% iou as compared to 46.0% secondISBI Cell tracking 2015, achieved 92% IoU as compared to 83%second

1Cite 9576 O. Ronneberger and P.Fischer and T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation,

Medical Image Computing and Computer-Assisted Intervention (MICCAI), LNCS-9351, pages 234–241, Springer-2015

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 3 / 8

Object Detection

What is there andwhere?

Deformative PartModel, andF-RCNN

Apply the model to an image at multiple locations and scales.High scoring regions are considered detections.

Yolo: apply a single neural network to the full image that divides it intoregions and predicts bounding boxes and probabilities for each region

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 4 / 8

Object Detection

What is there andwhere?

Deformative PartModel, andF-RCNN

Apply the model to an image at multiple locations and scales.High scoring regions are considered detections.

Yolo: apply a single neural network to the full image that divides it intoregions and predicts bounding boxes and probabilities for each region

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 4 / 8

Object Detection

What is there andwhere?

Deformative PartModel, andF-RCNN

Apply the model to an image at multiple locations and scales.High scoring regions are considered detections.

Yolo: apply a single neural network to the full image that divides it intoregions and predicts bounding boxes and probabilities for each region

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 4 / 8

Yolo 2

Conditional probability mapSee https://pjreddie.com/darknet/yolo/

2Cite 6375 Redmon, Joseph and Divvala, Santosh and Girshick, Ross and Farhadi, Ali, You only look once: Unified,

real-time object detection, IEEE conference on computer vision and pattern recognition (CVPR), pages 779–788, 2016

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 5 / 8

Yolo

S × S segments, gives B bounding boxes with confidence, and Cclass probabilities. So S × S × (B × 5 + C) values. S:7, B:2, C:20

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 6 / 8

Yolo

It is fastSpeed comes at the price of accuracy. Improved to 69%Generalizes wellLatest version YOLOv3 2018

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 7 / 8

Thank You!

Thank you very much for your attention3 !

Queries ?

3Credit: https://www.youtube.com/watch?v=NM6lrxy0bxs

STCS-DL4NLP&CV (CS-F441) Campus @ BITS-Pilani Lecture-KT-14 (Nov 20, 2019) 8 / 8