CoChat
Sign Up
Cited by 2Medical Image Segmentation TechniquesAdvanced Neural Network ApplicationsVisual Attention and Saliency Detection

Automatic Inside Point Localization with Deep Reinforcement Learning for Interactive Object Segmentation

Open access
Li, Guoqing, Zhang, Guoping, Qin, Chanchan·2021·Sensors
DOI 10.3390/s21186100PMC PMC8472885
Open this paper in CoChat — free accountOpen PDF

Abstract

In the task of interactive image segmentation, the Inside-Outside Guidance (IOG) algorithm has demonstrated superior segmentation performance leveraging Inside-Outside Guidance information. Nevertheless, we observe that the inconsistent input between training and testing when selecting the inside point will result in significant performance degradation. In this paper, a deep reinforcement learning framework, named Inside Point Localization Network (IPL-Net), is proposed to infer the suitable position for the inside point to help the IOG algorithm. Concretely, when a user first clicks two outside points at the symmetrical corner locations of the target object, our proposed system automatically generates the sequence of movement to localize the inside point. We then perform the IOG interactive segmentation method for precisely segmenting the target object of interest. The inside point localization problem is difficult to define as a supervised learning framework because it is expensive to collect image and their corresponding inside points. Therefore, we formulate this problem as Markov Decision Process (MDP) and then optimize it with Dueling Double Deep Q-Network (D3QN). We train our network on the PASCAL dataset and demonstrate that the network achieves excellent performance.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes IPL-Net, a deep reinforcement learning agent that automatically localizes the inside point for Inside-Outside Guidance (IOG) interactive object segmentation after a user clicks two outside points. It reports improved segmentation over outside-only baselines and reduces the required clicks from three to two.

Key findings

  • On PASCAL VOC val, IPL-Net outperforms the outside-only and two-stage methods by more than 0.8% and 0.6%, respectively (Table 1).
  • IPL-Net requires only two clicks to reach target IoU scores of 85% on PASCAL and 90% on GrabCut (Table 2).
  • Ablation: using the PASCAL_SBD pretrained model improves performance by 1% under PASCAL-1k training; with PASCAL-10k training, performance improves from 91.5% to 91.7% (Table 4).
  • The proposed reward combining IoU change and Centrality Degree performs better than using either reward alone, and initializing the inside point at the geometry center is significantly better than GrabCut center or random initialization (Figure 7).

Methods

  • The inside point localization problem is formulated as a Markov Decision Process: state is the cropped RGB image, the Gaussian map of the inside point, and the binary IOG segmentation mask; action is one of nine choices (up, down, left, right, four diagonals, still); reward is a weighted sum of IoU change and Centrality Degree change, with α = 0.7 and β = 0.3, plus a still-action reward of +2 or −2 depending on whether IoU is at least η.
  • The agent is trained with Dueling Double Deep Q-Network (D3QN), using Double DQN, Dueling DQN, and Prioritized Experience Replay. Input is five-channel 64×64 data; the network has three convolutional layers with 64, 128, and 256 filters, followed by a 256-unit fully connected layer. The movement step uses a = 15, and the process terminates on the still action or after 10 moves.
  • Training uses PASCAL VOC 2012: PASCAL-10k (10,582 images with 25,832 objects) and PASCAL-1k (1,464 images with 3,507 objects); evaluation uses the PASCAL val set (1,449 images with 3,427 objects). GrabCut has 50 natural scene images, and COCO contains objects in 80 categories. The model is trained for 50 epochs on PASCAL-1k or 30 epochs on PASCAL-10k, with minibatch size 512, Adam optimization, replay memory of 400,000 experiences, and ε annealed from 0.9 to 0.1 over the first 100,000 steps.

Limitations

  • The paper states that future work includes training the system with a more effective DRL algorithm to improve inside-point localization accuracy; it does not present a formal limitations section.
  • The text makes evident that the agent is trained on PASCAL, and although generalization is tested on GrabCut and COCO, IPL-Net shows “a little performance degradation” using only two clicks compared with the IOG result using the simulated inside point, which the paper calls an ideal value.
  • The ablation states that the geometry-center initialization is significantly better than GrabCut-center or random initialization because the maximum number of inside-point movements is 10, so other initial positions sometimes cannot reach the suitable position within 10 steps.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/s21186100

Cites

31 of 31 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

5 of 14 cited sources checked · 3 supported · 0 flagged · accuracy not yet measured

  • Findingnone given

    IPL-Net reduces the number of clicks required from three to two compared with interactive IOG segmentation.

    “Comparing with interactive IOG segmentation, we further reduce the number of clicks required from three to two.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For most objects, 10 steps is enough to localize the inside point that is initially located at the geometry center of the cropped image.
  • Findingown data

    In the Inside-Outside Guidance (IOG) interactive image segmentation algorithm, inconsistent selection of the inside point between training and testing causes significant performance degradation.

    “Nevertheless, we observe that the inconsistent input between training and testing when selecting the inside point will result in significant performance degradation.”
    • Also stated in: § 1. Introduction · hedged
      “However, we observe that the inconsistent choices in selecting the inside point between training and testing will often result in significant performance degradation.”
    • Also stated in: § 1. Introduction · plain
      “From Figure 1a, we can observe a large performance degradation when there is significant inconsistent selection between the annotations of the inside point and the simulated inside point.”
    Quote in the paperSubstantial claim
  • Findingown data

    The agent has the ability to find the optimal position of the inside point when there are unrelated objects lying inside the target object.

    “The agent also has the ability to find the optimal position of the inside point when there are unrelated objects lying inside the target object, as in the case of the five rows.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Segmentation performance is significantly improved when the inside point moves according to the proposed IPL-Net compared with the IOG segmentation corresponding to the initial inside point.

    “As shown in Figure 6, the results show that the segmentation performance is significantly improved when the inside point moves according to the proposed IPL-Net compared with the IOG segmentation corresponding to the initial inside point.”
    Quote in the paperSubstantial claim
  • Findingown data

    IPL-Net requires only two clicks to reach the target IoU accuracy of 85% on PASCAL and 90% on GrabCut.

    “As show in this table, we can see that our proposed IPL-Net requires only two clicks to reach the specific accuracy on both datasets.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 11
  • May combine two claims. 6
  • The quote only partly states this claim. 1

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Double DQN solves the overestimation problem commonly occurring in DQN.

    “The only difference was that Double DQN decoupled the selection from the evaluation to solve the overestimate problem that commonly occurs in DQN.”
    • Also stated in: § 2. Related Work · plain
      “The former was introduced to solve the over-estimation issue and yield accurate value estimates to improve the performance of the model”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [19] Hasselt, H.V., Guez, A., and Silver, D. (2016, January 12–17). Deep reinforcement learning with double Q-Learning. Proceedings of the National Conference on Artificial Intelligence, (AAAI-16), Phoenix, AZ, USA. source

      The Double DQN section (p. 4) says the algorithm separates action selection from evaluation to reduce overestimation, but explicitly says they are “not fully decoupled.” The Results on overoptimism (p. 5) report that DQN overestimated values in all 49 tested Atari games and that Double DQN’s estimates were closer to the true values. The paper supports reducing the problem, not solving it completely. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Dueling DQN framework separates the representation of state values and action advantages.

    “we adopt the framework of the Dueling DQN [20], where the representation of state values and action advantages are separated.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] Wang, Z., Schaul, T., Hessel, M., Hasselt, H.V., Lanctot, M., and Freitas, N.D. (2016, January 19–24). Dueling network architectures for deep reinforcement learning. Proceedings of the International Conference on Machine Learning, New York City, NY, USA. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Dueling DQN learns the state-value function efficiently by separating the representation of state values and action advantages.

    “while the latter was used for the purpose of learning the state-value function efficiently, where the representation of state values and action advantages are separated.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] Wang, Z., Schaul, T., Hessel, M., Hasselt, H.V., Lanctot, M., and Freitas, N.D. (2016, January 19–24). Dueling network architectures for deep reinforcement learning. Proceedings of the International Conference on Machine Learning, New York City, NY, USA. source

  • Classified as probably not a substantial claim.

    Premisecitation

    Dueling Double Deep Q-Network (D3QN) copes well with discrete action spaces.

    “The agent was trained using Dueling Double Deep Q Network (D3QN) [18,19,20], which copes well with the discrete action spaces.”
    • Also stated in: § 3.1. Overview · boosted
      “We formulate this problem as a Markov Decision Process (MDP) and then optimize it by reinforcement learning paradigm, more specifically Dueling Double Deep Q Network (D3QN), which copes well with the discrete action space of the agent.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [18] Mnih, 2015, Human-level control through deep reinforcement learning, Nature source
    • cites [19] Hasselt, H.V., Guez, A., and Silver, D. (2016, January 12–17). Deep reinforcement learning with double Q-Learning. Proceedings of the National Conference on Artificial Intelligence, (AAAI-16), Phoenix, AZ, USA. source

      The Deep Q Networks section (p. 2) describes a network that outputs values for a finite set of actions, and Empirical results (p. 4) reports that Double DQN performs well on Atari games. The paper studies Double DQN, however, not the combined Dueling Double DQN (D3QN) named in the claim, so it does not establish that D3QN itself copes well with discrete action spaces. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [20] Wang, Z., Schaul, T., Hessel, M., Hasselt, H.V., Lanctot, M., and Freitas, N.D. (2016, January 19–24). Dueling network architectures for deep reinforcement learning. Proceedings of the International Conference on Machine Learning, New York City, NY, USA. source

  • May combine two claims.

    Findingown data

    The average IoU accuracy on the PASCAL val set reaches its maximum at the 50th epoch using the model trained on PASCAL-1k and at the 30th epoch using the model trained on PASCAL-10k.

    “From this Figure, we can see that the average IoU accuracy on PASCAL val set reaches the maximum at the 50th epoch using the model trained on PASCAL-1k and 30th epoch using the model trained on PASCAL-10k.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    It takes about 0.9 s to find the most suitable location of the inside point on a single NVIDIA 2080Ti GPU and 3.60 GHz Intel Core i9-9900K CPU.

    “It takes about 0.9 s to find the most suitable location of the inside point on a single NVIDIA 2080Ti GPU and 3.60 GHz Intel Core i9-9900K CPU (Xi’an Kunlong Computer Technology Co. LTD, Xi’an, China).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    IPL-Net outperforms the outside only and two-stage methods by more than 0.8% and 0.6% IoU accuracy on PASCAL val using outside points annotations only.

    “It can be observed that our IPL-Net outperforms the other two algorithm by more than 0.8% and 0.6%, which well validate the effectiveness of our proposed method in inferring the position of the inside point to help the IOG algorithm to achieve accurate performance.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16] Zhang, S., Liew, J.H., Wei, Y., Wei, S., and Zhao, Y. (2020, January 14–19). Interactive Object Segmentation with Inside-Outside Guidance. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. source

  • The quote only partly states this claim.

    Findingown data

    IPL-Net achieves a little performance degradation when using only two clicks on COCO MVal, COCO MVal seen, COCO MVal unseen, and GrabCut compared with the IOG algorithm using a simulated inside point.

    “It can be observed that our IPL-Net achieves a little performance degradation when using only two clicks.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The IOG segmentation result obtained with the simulated inside point is an ideal value.

  • Classified as probably not a substantial claim.

    Premisecitation

    The IOG algorithm chooses the point whose ε(p)=1 in cropped images as the simulated inside point.

    “The IOG algorithm chooses the point whose ε(p)=1 in cropped images as the simulated inside point.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16] Zhang, S., Liew, J.H., Wei, Y., Wei, S., and Zhao, Y. (2020, January 14–19). Interactive Object Segmentation with Inside-Outside Guidance. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. source

  • May combine two claims.

    Premisecitation

    Applying Prioritized Experience Replay (PER) boosts sample efficiency and improves the performance of the agent.

    “During training, we also apply the trick of Prioritized Experience Replay (PER) [31] to boost the sample efficiency and improve the performance for the agent.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] Schaul, T., Quan, J., Antonoglou, I., and Silver, D. (2016, January 2–4). Prioritized Experience Replay. Proceedings of the International Conference on Learning Representations, San Juan, Puerto Rico. source

  • Classified as probably not a substantial claim.

    Premisecitation

    In IOG training, the position of the object center is simulated as the inside point.

    “According to IOG algorithm, the position of the object center is simulated as the inside point during training.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16] Zhang, S., Liew, J.H., Wei, Y., Wei, S., and Zhao, Y. (2020, January 14–19). Interactive Object Segmentation with Inside-Outside Guidance. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The Inside-Outside Guidance (IOG) algorithm achieves fast clicking speed and excellent segmentation results across different application domains.

    “Despite its simplicity, the Inside-Outside Guidance (IOG) algorithm has shown fast clicking speed and excellent segmentation results across different application domains.”
    • Also stated in: § 2. Related Work · plain
      “Despite its simplicity, the IOG algorithm has achieved high performance across different application domains.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [16] Zhang, S., Liew, J.H., Wei, Y., Wei, S., and Zhao, Y. (2020, January 14–19). Interactive Object Segmentation with Inside-Outside Guidance. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. source

  • May combine two claims.

    Interpretationown data

    The actions chosen by the agent attempt to move the inside point close to the simulated one and to help the interactive IOG segmentation algorithm obtain accuracy in the segmentation result.

    “Notice that the actions chosen attempt to move the inside point to close to the simulated one, and also to help the interactive IOG segmentation algorithm to obtain accuracy in the segmentation result.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The simulated inside point used as the target for training the agent is the most suitable position for the inside point.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Prioritized Experience Replay replaces DQN's uniform sampling strategy from the replay memory with a novel sampling strategy to accelerate learning dramatically.

    “In addition, we also use the concept of Prioritized Experience Replay [31], which replaced DQN’s uniform sampling strategy from the replay memory with a novel sampling strategy to accelerate learning dramatically.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] Schaul, T., Quan, J., Antonoglou, I., and Silver, D. (2016, January 2–4). Prioritized Experience Replay. Proceedings of the International Conference on Learning Representations, San Juan, Puerto Rico. source

  • Classified as probably not a substantial claim.

    Premisecitation

    The Inside-Outside Guidance (IOG) algorithm requires three points: one inside point located around the center of the object and two outside points at symmetrical corner locations that form an almost-tight bounding box enclosing the target object.

    “In an IOG algorithm, the workload of the annotator is reduced to only providing three points: an inside point located around the center of the object and two outside points at any symmetrical corner locations that form an almost-tight bounding box enclosing the target object of interest.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16] Zhang, S., Liew, J.H., Wei, Y., Wei, S., and Zhao, Y. (2020, January 14–19). Interactive Object Segmentation with Inside-Outside Guidance. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. source

  • May combine two claims.

    Interpretationreasoning

    The proposed agent learnt from its own mistakes and optimized the policy to find the most suitable position for the inside point.

    “The proposed agent learnt from its own mistakes and optimized the policy to find the most suitable position for the inside point.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    It is meaningful that IPL-Net can locate the inside point automatically.

    “Therefore, it is meaningful that our IPL-Net can locate the inside point automatically.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The IOG segmentation result obtained with the simulated inside point is an ideal value.
17 of 17

  • The paper relies on this without support.

    Collecting images and their corresponding inside points is expensive, making the inside point localization problem difficult to define as a supervised learning framework.

    “The inside point localization problem is difficult to define as a supervised learning framework because it is expensive to collect image and their corresponding inside points.”
    “In this study, the problem is difficult to define as a supervised learning framework, because it is expensive to collect images and their corresponding inside points.”

  • The authors state this as an assumption.

    For most objects, 10 steps is enough to localize the inside point that is initially located at the geometry center of the cropped image.

    “We believe that for most objects, 10 steps is enough to localize the position of the inside point that is initially located at the geometry center of the cropped image.”

  • The paper relies on this without support.

    Defining the action corresponding to all pixels in the cropped image makes the action space too large, leading to training problems.

    “If we define the action corresponding to all the pixels in the cropped image, the action space becomes too large, leading to training problems.”

  • The paper relies on this without support.

    The IOG segmentation result obtained with the simulated inside point is an ideal value.

    “It is to be noticed that the result of the IOG is an ideal value.”

  • The paper relies on this without support.

    Collecting images and their corresponding inside points is high cost, making it difficult to define the inside point localization problem as a supervised learning framework.

    “Due to the high cost of collecting the images and their corresponding inside points, it is difficult to define this problem as a supervised leaning framework.”

  • The paper relies on this without support.

    The simulated inside point used as the target for training the agent is the most suitable position for the inside point.

    “Notice that the actions chosen attempt to move the inside point to close to the simulated one”
6 of 6

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Reference [9] has conflicting author attribution

    Minor

    The text attributes reference [9] to Xu et al., while the listed reference begins with Ning X. and does not list Xu.

  • Adam optimizer citation points to a segmentation paper

    Minor

    The training section cites [9] for the Adam optimizer, but reference [9] is listed as "Deep Interactive Object Selection," not a paper on Adam.

  • Gaussian inside-point map is also called binary

    Minor

    The state definition says both the inside-point map and segmentation mask are binary images, but elsewhere describes the inside-point map as Gaussian-encoded. A Gaussian map is not a binary image.

3 of 3

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Ning, X., Price, B., Cohen, S., Yang, J., and Huang, T. (2016, January 27–30). Deep Interactive Object Selection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA.

Cited by

Sorted by citations

Related

Authors
Li, G., Zhang, G., & Qin, C.
Affiliations
  • Central China Normal University (CN)
  • Guizhou Normal University (CN)
Corresponding author
Qin, Chanchan
ORCID
Venue
Sensors
Publisher
MDPI AG
Published
2021-09-11
Volume
21
Issue
18
Pages
6100
ISSN
1424-8220 (ISSN-L)
DOI
10.3390/s21186100
PMC
PMC8472885
OpenAlex
W3199678105
Semantic Scholar
ddd9d6fdbd4d8824dfdeedacbfdcad22181924e8
Cited by
2
References
29
Citations by year
2025: 1 · 2024: 1
FWCI
0.09
Type
journal article
Language
en
Open access
Licence
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Europe PMC (PubMed Central) (submittedVersion, other-oa)
  • PubMed Central (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Natural Science Foundation of Hubei Province (2017CFB591)
  • Natural Science Research Foundation of Education Department of Guizhou Province ([2021]305)
  • Doctoral Scientific Research Foundation of Guizhou Normal University (2014)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Visual Attention and Saliency Detection
Keywords
interactive segmentation, deep reinforcement learning, Markov decision process, dueling double deep Q-network, user-guided segmentation, PASCAL dataset
MeSH terms
Algorithms, Image Processing, Computer-Assisted

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-21.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-21.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review