본 논문에서는 interactive segmentation 문제를 풀기 위하여 deep reinforcement learning을 활용한 seed gereration 기법을 제안한다. Interactive segmentation 문제의 이슈 중 하나는 사용자의 개입을 최소화하는 것이다. 본 논문에서 제안하는 시스템이 사용자를 대신하여 인공적인 seed를 생성하게 된다. 사용자는 initial seed 정보만을 제공하면 된다. 우리는 optimal seed point 정의의 모호함으로 인해 supervised 기법을 사용하여 학습하기 어려운 점을 reinforcement learning 기법을 사용하여 극복하였다. Seed generation 문제에 맞도록 MDP를 정의하여 deep-q-network를 성공적으로 학습하였다. 우리는 MSRA10K 데이터셋에 대하여 학습을 진행하여 기존 segmentation 알고리즘의 부정확한 initial 결과 대비 우수한 성능을 보였다.
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
Seed net automatic seed generation with deep reinforcement learning for robust interactive segmentation
1. SeedNet : Automatic Seed Generation with Deep
Reinforcement Learning for Robust Interactive
Segmentation
Gwangmo Song
Dept. of ECE, Seoul National University
2. Goal : Interactive Segmentation
Segmentation
Cutting out a desired object from the image
3. Goal : Interactive Segmentation
Interactive Segmentation
Cutting out a desired object from the image using user input
Camel!
7. Previous Approach
Early approaches : Graph Cut, GrabCut, Random Walk, Random
Walk with Restart, Geodesic [1,2,3,4,5]
Using MRF optimization to extract region of interest
Determine the region by comparing the correspondence between labeled and
unlabeled pixels
[1] Y. Y. Boykov et al. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In ICCV, 2001.
[2] C. Rother et al. Grabcut: Interactive foreground extraction using iterated graph cuts. In ToG. 2004.
[3] L. Grady. Random walks for image segmentation. In PAMI. 2006.
[4] T. H. Kim et al. Generative image segmentation using random walks with restart. In ECCV. 2008.
[5] V. Gulshan et al. Geodesic star convexity for interactive image segmentation. In CVPR. IEEE, 2010.
8. Previous Approach
Recent approaches : Deep Learning [1,2] …
Fine-tuning semantic segmentation network (FCN [3]) for interactive
segmentation task
End-to-end training by concatenating seed point and image
Significant performance improvements compared to classical
techniques
[1] N. Xu et al. Deep interactive object selection. In CVPR. 2016.
[2] J. Hao Liew et al. Regional interactive image segmentation networks. In ICCV. 2017.
[3] J. Long et al. Fully convolutional networks for semantic segmentation. In CVPR. 2015
9. Challenges
Minimizing user interaction
Main metric of segmentation : Accuracy %, Click #
Reducing the input burden of the user
Deep learning frameworks still require a large number of clicks
Segmentation Pascal Grabcut Berkeley
MSCOCO
(seen)
MSCOCO
(unseen)
Graph Cut [1] 15.06 11.10 14.33 18.67 17.80
Random Walk [2] 11.37 12.30 14.02 13.91 11.53
Geodesic [3] 11.73 8.38 12.57 14.37 12.45
iFCN [4] 6.88 6.04 8.65 8.31 7.82
RIS-Net [5] 5.12 5.00 6.03 5.98 6.44
[1] Y. Y. Boykov et al. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In ICCV, 2001.
[2] L. Grady. Random walks for image segmentation. In PAMI. 2006.
[3] V. Gulshan et al. Geodesic star convexity for interactive image segmentation. In CVPR. IEEE, 2010.
[4] N. Xu et al. Deep interactive object selection. In CVPR. 2016.
[5] J. Hao Liew et al. Regional interactive image segmentation networks. In ICCV. 2017.
10. Motivation
Expansion of seed information
Assuming a minimal user input (ex. Single seed point)
Automatically expand seed information and perform segmentation
Initial Seed RWR result Expanded Seed RWR result
11. Motivation
Problems of previous works
Existing works are only interested in segmentation accuracy with static
user input
Even SOTA gives unsatisfactory result with fewer user input
Any model can give good result with sufficient user input
Segmentation
Module
12. Motivation
Automatic seed control
Generate enough seed information from sparse input
Additional automatic seed control module
Combine with segmentation to configure the entire system
How to design seed control module?
Seed Control
Module
Segmentation
Module
15. Our Previous Work
Interactive Segmentation with Seed Expansion
Seed Expansion (Two stages)
• Superpixel-based
• Pixel-level to SP-level
• RWR & threshold
• Add pixels with distinct properties
to the seed
• Removes ambiguous pixels
through seed erase step
16. Our Previous Work
Interactive Segmentation with Seed Expansion
Image Pyramid
• Speed up
• Matrix inversion
(Coarsest level)
• Power iteration
(Refinement)
17. Our Previous Work
Interactive Segmentation with Seed Expansion
Refinement
• Saliency-based
18. Our Previous Work
Results
Limitation
Heuristic approach
Insufficient accuracy
Method F-score (%)
RWR [1] 74.48
Ours 80.32
[1] T. H. Kim et al. Generative image segmentation using random walks with restart. In ECCV. 2008.
Image &
Scribble
RWR
Results
Expanded
Seed
Our
Results
19. Our Approach
Automatic seed generation system with deep reinforcement
learning
Create additional seed using reinforcement learning
Simulate the human process
21. Why RL?
Learning framework
Supervised : label
Unsupervised : no label
Reinforcement : reward
RL
Decision making
In a state, agent does an action, get a reward
Supervised
Unsupervised Reinforcement
Learning
22. Why RL?
Optimal seed point
Differs by person
Depends on the current and future state
Person 1 Person 2 Person 3 Person 4
23. Why RL?
RL properties
Use reward instead of label that is hard to define
Instant reward + future reward
Optimized for sequential decision making problems (Atari games)
Interpret seed generation problem as seed locating game
24. RL Framework
Markov Decision Process
State
Action
Reward
Agent
Environment
ActionState Reward
25. RL Framework
Markov Decision Process
State : Input image, segmentation mask
Action : New seed point location
Reward : IoU(Intersection over Union)-based function
Seed
Generator
Segmentation
Module
New seed
location
Image
/ Mask
IoU
Score
26. Proposed Model
Propose seed generation network
Combine with existing segmentation module to configure the entire system
DQN
SEED
GENERATION
SEED
UPDATE
IMAGESEEDGTMASK
SEGMENTATION
SEGMENTATION
REWARD
OBSERVATION UPDATE
REWARD COMPUTATION
28. SEED
GENERATION
SEED
UPDATE
SEED
GENERATION
Action
Positioning new seed point on 2D grid
20×20 grid with 2 label (fore, back) = 800 action space
DQN
SEED
GENERATION
SEED
UPDATE
IMAGESEEDGTMASK
SEGMENTATION
SEGMENTATION
REWARD
OBSERVATION UPDATE
REWARD COMPUTATION
SEED
29. Criteria for updating the network for the action and state
Based on IoU score
REWARD COMPUTATION
GTMASK
SEGMENTATION
REWARD
REWARD COMPUTATION
Reward
DQN
SEED
GENERATION
SEED
UPDATE
IMAGESEEDGTMASK
SEGMENTATION
SEGMENTATION
REWARD
OBSERVATION UPDATE
REWARD COMPUTATION
SEGMENTATION
36. DQN
Train seed generation agent with RL
Deep Q-Network
Policy Gradient
Actor-Critic
Deep Q-Network
Train action-value function 𝑄(𝑎, 𝑠), the expected reward
Approximates 𝑄(𝑎, 𝑠) with a deep neural network
Loss function
𝐿𝑜𝑠𝑠 𝜃 = 𝔼 𝑟 + 𝛾 max
𝑎′
𝑄 𝑠′, 𝑎′; 𝜃 − 𝑄 𝑠, 𝑎; 𝜃
2
37. DQN
DQN architecture
Basic architecture from [1]
Double DQN [2]
Dueling DQN [3]
[1] V. Mnih et al. Human-level control through deep reinforcement learning. In Nature. 2015.
[2] H. Van Hasselt et al. Deep reinforcement learning with double q-learning. In AAAI, 2016.
[3] Z. Wang et al. Dueling network architectures for deep reinforcement learning. In ICML, 2016.
38. Experiments
MSRA10K dataset
Saliency task
10,000 images (9,000 train, 1,000 test)
Initial seed generated from GT mask
1 foreground, 1 background seed point
Our model
Segmentation module : Random Walk [1]
Stop after 10 seed generation
[1] L. Grady. Random walks for image segmentation. In PAMI. 2006.
39. MSRA10K
Quantitative Results
5 random seed sets
Intersection-over-Union score comparison
Method Set 1 Set 2 Set 3 Set 4 Set 5 Mean
RW [1] 39.59 39.65 39.71 39.77 39.89 39.72
Ours 60.70 60.12 61.28 61.87 60.90 60.97
[1] L. Grady. Random walks for image segmentation. In PAMI. 2006.
40. MSRA10K
Image GT RW results [1]
(Initial Seed)
Step 1 Step 2 Step 3 Step 10
Our results
[1] L. Grady. Random walks for image segmentation. In PAMI. 2006.
41. MSRA10K
Comparison with supervised methods
Same condition (Network configuration, Scratch)
Directly output segmentation mask
Input type : image only (FCN), with seed (iFCN)
Method FCN [1] iFCN [2] Ours
IoU 37.20 44.60 60.97
[1] J. Long et al. Fully convolutional networks for semantic segmentation. In CVPR. 2015
[2] N. Xu et al. Deep interactive object selection. In CVPR. 2016.
42. Ablation Study
Reward function
Method RW [1] 𝑹 𝑰𝒐𝑼 𝑹 𝒅𝒊𝒇𝒇 𝑹 𝒔𝒊
IoU 39.72 42.55 44.45 60.97
𝑅𝐼𝑜𝑈 𝑅 𝑑𝑖𝑓𝑓 𝑅 𝑠𝑖
[1] L. Grady. Random walks for image segmentation. In PAMI. 2006.
43. Ablation Study
Other segmentation module
Method GC [1] Ours (GC) GSC [2] Ours (GSC)
IoU 38.44 52.10 58.34 63.48
RWR [3] Ours (RWR)
35.71 53.04
Image GT Initial Ours
GCGSCRWR
[1] C. Rother et al. Grabcut: Interactive foreground extraction using iterated graph cuts. In ToG. 2004.
[2] V. Gulshan et al. Geodesic star convexity for interactive image segmentation. In CVPR. IEEE, 2010.
[3] T. H. Kim et al. Generative image segmentation using random walks with restart. In ECCV. 2008.
44. Unseen Dataset
Train on MSRA10K
Test on
GSCSEQ, Weizmann Single Object, Weizmann Horse, iCoseg, IG02
Dataset GSCSEQ
Weizmann
Single
Weizmann
Horse
iCoseg IG02
Method RW Ours RW Ours RW Ours RW Ours RW Ours
IoU 27.40 38.93 38.29 55.08 25.63 48.33 36.02 43.87 23.87 28.48
47. Summary
Suggest novel seed generation module with deep reinforcement
learning
Improved segmentation accuracy using automatic seed
generation system
Validation on various segmentation modules and datasets