Tweet
Tweet

Slide 1

Slide 1 text

Jumabek Alikhanov @Information Security Research Lab, Inha University YOLO9000: Better, Faster, Stronger (CVPR 2017, Best Paper Honorable Mention) 1

Slide 2

Slide 2 text

1. Introduction & Previous Work 2. Better detection performance 3. Faster processing speed 4. Detecting more classes(object types) 5. Conclusion CONTENTS 2

Slide 3

Slide 3 text

Task & Evaluation Metric mAP- mean Avarage Precision 3 https://github.com/rafaelpadilla/Object-Detection-Metrics

Slide 4

Slide 4 text

YOLO v1 Network Output shape = (S, S, B×5 + C) = (7, 7, 2×5 + 20) = (7, 7, 30). 4

Slide 5

Slide 5 text

YOLOv1: Loss Function pi-conditional class Prob. Ci - box confidence score 5 Localization Confidence Classification

Slide 6

Slide 6 text

Previously Pascal 2007 mAP Speed DPM v5 33.7 .07 FPS 14 s/img R-CNN 66.0 .05 FPS 20 s/img Fast R-CNN 70.0 .5 FPS 2 s/img Faster R-CNN 73.2 7 FPS 140 ms/img YOLO 63.4 45 FPS 22 ms/img 6

Slide 7

Slide 7 text

Previously Pascal 2007 mAP Speed DPM v5 33.7 .07 FPS 14 s/img R-CNN 66.0 .05 FPS 20 s/img Fast R-CNN 70.0 .5 FPS 2 s/img Faster R-CNN 73.2 7 FPS 140 ms/img YOLO 63.4 45 FPS 22 ms/img 7

Slide 8

Slide 8 text

Better Performance 8

Slide 9

Slide 9 text

9 YOLO Train on ImageNet Fine-tune on detection Resize network

Slide 10

Slide 10 text

10 Fine-tune 448x448 Classifier: +3.5% mAP Train on ImageNet Fine-tune on detection Resize, fine-tune on ImageNet

Slide 11

Slide 11 text

Anchor boxes use static initialization

Slide 12

Slide 12 text

Use k-means clustering to find better initializations https://github.com/Jumabek/darknet_scripts

Slide 13

Slide 13 text

No content

Slide 14

Slide 14 text

Static Anchors vs Dimension Clusters 14

Slide 15

Slide 15 text

Box Location Prediction 15

Slide 16

Slide 16 text

Dimension Clusters: +5% mAP

Slide 17

Slide 17 text

17 Multi-scale training: +1.5% mAP

Slide 18

Slide 18 text

YOLOv2: Fast, Accurate Detection

Slide 19

Slide 19 text

Huang, Jonathan, et al. "Speed/accuracy trade-offs for modern convolutional object detectors." arXiv preprint arXiv:1611.10012 (2016).

Slide 20

Slide 20 text

Huang, Jonathan, et al. "Speed/accuracy trade-offs for modern convolutional object detectors." arXiv preprint arXiv:1611.10012 (2016).

Slide 21

Slide 21 text

Huang, Jonathan, et al. "Speed/accuracy trade-offs for modern convolutional object detectors." arXiv preprint arXiv:1611.10012 (2016). YOLOv2

Slide 22

Slide 22 text

No content

Slide 23

Slide 23 text

Faster Detection Speed 23

Slide 24

Slide 24 text

Speed is not just parameter counts or FLOPs Top 1 Top 5 FLOPs GPU Speed VGG-16 70.5 90.0 30.95 Bn 100 FPS Extraction (YOLOv1) 72.5 90.8 8.52 Bn 180 FPS Resnet50 75.3 92.2 7.66 Bn 90 FPS

Slide 25

Slide 25 text

Darknet19: A good balance of speed and accuracy Top 1 Top 5 FLOPs GPU Speed VGG-16 70.5 90.0 30.95 Bn 100 FPS Extraction (YOLOv1) 72.5 90.8 8.52 Bn 180 FPS Resnet50 75.3 92.2 7.66 Bn 90 FPS Darknet19 74.0 91.8 5.58 Bn 200 FPS

Slide 26

Slide 26 text

Why is it fast? Simple & efficient architecture C implementation 26

Slide 27

Slide 27 text

Stronger - Detecting more classes 27

Slide 28

Slide 28 text

- 14 million images - 22k classes - Classification labels - 100k images - 80 classes - Detection labels Golden eagle

Slide 29

Slide 29 text

Typically use softmax over all classes

Slide 30

Slide 30 text

Can’t just mash classes together...

Slide 31

Slide 31 text

Can’t just mash classes together...

Slide 32

Slide 32 text

WordNet has structure but it’s messy

Slide 33

Slide 33 text

No content

Slide 34

Slide 34 text

No content

Slide 35

Slide 35 text

... Each node is a conditional probability

Slide 36

Slide 36 text

... Each node is a conditional probability P(Bedlington terrier) = P(object) * P (living thing | object) * ….. P(canine | mammal) * P(dog | canine) * P(terrier | dog) * P(Bedlington terrier | terrier)

Slide 37

Slide 37 text

No content

Slide 38

Slide 38 text

No content

Slide 39

Slide 39 text

No content

Slide 40

Slide 40 text

No content

Slide 41

Slide 41 text

No content

Slide 42

Slide 42 text

No content

Slide 43

Slide 43 text

No content

Slide 44

Slide 44 text

No content

Slide 45

Slide 45 text

No content

Slide 46

Slide 46 text

No content

Slide 47

Slide 47 text

Conclusion ● YOLOv2 and YOLO9000 real-time detection systems ● YOLOv2 state of the art and faster than other systems ● 9K object category detection by YOLO9000 47

Slide 48

Slide 48 text

1. CVPR paper - https://pjreddie.com/media/files/papers/YOLO9000.pdf 2. Article - https://medium.com/@jonathan_hui/real-time-object-detection-with-yolo-yolov2-28b1b93e2088 3. Author’s Presentation - https://docs.google.com/presentation/d/14qBAiyhMOFl_wZW4dA1CkixgXwf0zKGbpw_0oHK8yEM/edit#slide=id.g1f9fb98e4b_0 _132 References 48