Pith. sign in

REVIEW 3 major objections 6 minor 34 references

VILOD: A Visual Interactive Labeling Tool for Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This thesis claims that a balanced, visually guided human labeling strategy in VILOD achieves higher final object detection performance than automated uncertainty sampling, and that visual analytics can make active learning more transparent

desk verdict A transparent master's thesis that builds a plausible visual active-learning tool for object detection; the central performance claim is a proof-of-concept, not a demonstrated edge. read the letter →

arxiv 2509.05317 v1 pith:ZMSBEDSF submitted 2025-08-29 cs.CV cs.AIcs.HCcs.LG

Isac Holm This is my paper · ORCID
classification cs.CVcs.AIcs.HCcs.LG
keywords ObjectDetectionActiveLearningHuman-in-the-LoopVisualAnalyticsInteractiveLabelingt-SNEUncertaintySamplingTransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This master's thesis develops VILOD, an interactive visual labeling tool for object detection, and tests whether expert users guided by visual analytics can label data as effectively as automated active learning. The central claim, based on a comparative study of three labeling strategies, is that a "Balanced Guidance Integration" strategy—using all available cues: data structure, uncertainty heatmaps, active-learning suggestions, and model state—achieves the highest final object detection performance, slightly exceeding an automated uncertainty-sampling baseline. The paper also argues that the tool makes model state and dataset characteristics interpretable enough for experts to implement distinct strategies, and that human quality control prevents the performance dips that purely automated sampling can suffer. If correct, this supports human-in-the-loop active learning as a competitive and more transparent alternative to fully automated annotation for object detection.

What carries the argument

The central mechanism is the VILOD interface: a t-SNE scatterplot of image features forms the Data View, with an uncertainty heatmap (a kernel density estimate weighted by squared uncertainty) overlaid to highlight regions of high model uncertainty, and active-learning-suggested samples marked distinctly. A linked Model View shows per-class prediction-confidence distributions and the class balance of the labeled set. Together with a lasso selection tool and an integrated bounding-box annotation modal, these views let the user choose samples according to different strategies and immediately retrain the detector, closing a human-in-the-loop active learning cycle.

What would settle it

Run the same three strategies and the uncertainty baseline with several independent annotators (not the author) and multiple random seeds on the same dataset; if the balanced strategy does not reliably reach the highest final mAP, or if the differences fall within seed-to-seed variance, the claim of a human-guidance advantage fails.

Watch

Extended reading notes

Core claim

The paper's core discovery claim is that an expert user, supported by an interactive visual analytics interface, can implement distinct labeling strategies for object detection annotation, and that one strategy in particular—Balanced Guidance Integration, which synthesizes uncertainty heatmaps, active-learning suggestions, t-SNE data structure, and model-state feedback—yields the highest final model performance. In the reported comparison, this balanced human-guided strategy reached the best mAP50-95, slightly outperforming the automated uncertainty-sampling active learning baseline, while the two more narrow strategies (exploration-only and uncertainty-only) did not surpass the baseline on

Load-bearing premise

The comparison rests on one person—the author—performing the three strategies, so the measured edge of the balanced strategy could reflect that person's skill, expectations, or fatigue rather than the strategy itself; a single training seed adds to this fragility.

Editorial extensions

If this is right

  • If the central claim holds, human-in-the-loop visual analytics can be a viable alternative to fully automated active learning for object detection, offering comparable or better performance with added transparency.
  • Experts can pursue diverse labeling goals—broad data coverage, uncertainty reduction, class balancing, or correcting specific model errors—within one integrated tool, rather than being locked into a single query strategy.
  • Human quality control can filter out noisy, low-resolution, or misleading samples that a purely automated uncertainty sampler might add, potentially avoiding performance dips during iterative training.
  • The tool makes the model's state interpretable, so practitioners can see why samples are suggested and decide whether to trust or override algorithmic recommendations.
  • The comparison of strategies provides empirical grounding for designing adaptive or hybrid active learning interfaces that encourage balanced use of data structure and model uncertainty.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If replicated with multiple independent annotators and multiple random seeds, the balanced-strategy advantage may become more consistent; a multi-user study could separate strategy effects from the single author's familiarity with the tool.
  • The t-SNE structure comes from a fixed pretrained feature extractor; updating the projection from each newly trained model's embeddings could make the Data View model-aware and potentially improve later iterations, an extension the paper notes as future work.
  • The balanced strategy could be partially automated by combining uncertainty scores with a diversity/representativeness term, turning the expert's balancing act into a hybrid query strategy while keeping the human as final filter.
  • The quality-control benefit observed in the uncertainty-driven case suggests that even users who largely follow active learning suggestions add value mainly by discarding noisy samples, a finding that could be tested by comparing an AL baseline with and without a simple automated noise filter.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper presents VILOD, a web-based visual interactive labeling tool for object detection that combines a t-SNE scatterplot, uncertainty heatmap overlay, active-learning (AL) suggestions, model-state views, and an integrated bounding-box annotation interface. It reports a comparative study in which the author, acting as an expert user, executed three labeling strategies (Exploration & Structure Focus, Uncertainty-Driven Focus, Balanced Guidance Integration) over five iterative HITL-AL rounds on the African Wildlife dataset with YOLOv11n, and compared their mAP50-95 trajectories against an automated uncertainty-sampling baseline. The central claim is that the Balanced Guidance Integration strategy achieved the highest final mAP50-95 and that human guidance supported by interactive visual analytics can offer an edge over automated AL.

Significance. If the empirical claim were supported, the paper would make a useful contribution to HITL active learning for object detection by demonstrating a concrete system and showing that different visually guided strategies yield measurable differences in learning trajectories. The tool itself is thoughtfully designed and described in detail, and the qualitative walkthroughs (Sections 4.1–4.3) are valuable for the visual-analytics community as proof-of-concept evidence. The paper is also unusually transparent about its limitations, explicitly acknowledging the single-user design, single dataset, single model, and lack of statistical testing in Sections 1.6, 3.5, and 6.2. However, the central claim of a performance edge over automated AL rests on a single replication with point estimates only, and the paper does not provide the variance estimates or repeated runs needed to support that causal reading.

major comments (3)
  1. [§4.4.1, Table 4.2] The headline result — Balanced Guidance Integration (0.7477) outperforming the automated AL baseline (0.7348) — is based on one seed (seed=42, §3.3.2), one dataset, one user, and no confidence intervals or significance tests. The 0.0129 mAP difference is smaller than the iteration-to-iteration variation in the same table (e.g., AL baseline drops from 0.6940 to 0.6824 between iterations 2 and 3). Without repeated training runs with different seeds or multiple users, the difference cannot be distinguished from stochasticity or selection variance. The abstract and §6.1 present the edge claim without this qualification; this is load-bearing for RQ2.
  2. [§3.3.7, §3.5] The three use cases were executed by the author only, as the paper states in §3.5. The qualitative narratives in Sections 4.1–4.3 describe strategic decisions in a way that supports the intended strategy, but there is no independent protocol check, no rater agreement, and no log-based measure of how faithfully the stated strategy was followed. The causal inference that observed mAP differences reflect the strategies rather than the user's expectations or fatigue is therefore weak. The paper acknowledges this limitation but still concludes in §6.1 that the system 'successfully demonstrated' benefits; the conclusion should be explicitly framed as a proof-of-concept result.
  3. [§4.4.1, iteration-3 AL baseline] The sudden drop in the AL baseline trajectory at iteration 3 is attributed to 'noisy samples' that a human might filter. This is a plausible post-hoc interpretation, but it is not supported by any measurement: the paper does not analyze which images were selected at that iteration, does not quantify their noise content, and does not compare them with images selected in other iterations. Since this interpretation is used to motivate the value of human quality control, it should be either substantiated with evidence or presented as speculation.
minor comments (6)
  1. [§3.3.2] The checkpoint is referred to as 'yolov11n.pt' and elsewhere as 'YOLOv11n'; please standardize the naming.
  2. [§3.3.4, Figure 3.6] The caption labels two components as '(c)' — the Selected Images View and the Labeled View. Renumber.
  3. [§4.4.1, Table 4.3] Precision and Recall are reported without defining the confidence threshold or IoU setting; please specify these evaluation details.
  4. [§4.1.1] The phrase 'the three blobs' is clear from Figure 4.16, but the figure is a screenshot with manual highlighting; a vectorized figure would improve readability.
  5. [§6.2] The sentence 'investigation if there is any significant performance increase' should be 'investigate whether there is a significant performance increase.'
  6. [References] Some references lack venue identifiers (e.g., [13], [31]); consider adding DOIs or arXiv identifiers where available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the mAP comparison is measured on a held-out test set with no fitted parameter or self-cited theorem manufacturing the result.

full rationale

The paper's central claim is that the Balanced Guidance Integration strategy achieved the highest final mAP50-95 (0.7477) among the compared strategies, slightly surpassing the automated AL baseline (0.7348). This is an empirical result obtained by running three predefined human-guided use cases and one automated simulation, all starting from the same initial model M0, retraining for five iterations with the same fixed hyperparameters (epochs=50, seed=42, imgsz=640), and evaluating on a held-out test split of 227 images that was excluded from all selection and training processes. No equation in the paper defines the strategy's performance in terms of its inputs; the mAP numbers come from model evaluation, not from a fitted parameter renamed as a prediction. The AL suggestions and uncertainty heatmap are computed directly from the current model's predictions, and the human strategies are described as distinct selection policies over those visualizations. There is no self-citation that is load-bearing for the main claim: references to the supervisor's prior work appear only in background and related-work contexts (e.g., trajectory annotation, visual analytics platforms) and do not establish the empirical outcome. The paper transparently acknowledges threats to validity—single user, single dataset, single seed, lack of significance testing—in Sections 1.6, 3.5, 5, and 6.2. These are limitations on generalizability and statistical inference, not definitional circularity. The evaluation loop (model predicts, user selects, model retrains, test on held-out set) is the standard active learning protocol; nothing in the paper reduces the claimed performance advantage to its own construction. Therefore, no circular step is present, and the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The comparison depends on several hand-picked hyperparameters (t-SNE perplexity, K, heatmap exponent, budget, epochs, seed) and on domain assumptions about the validity of the visual representation and the fidelity of the single expert user. None of these are fit to the held-out test set, so the circularity burden is low, but the free-parameter count shows that the reported mAP differences are not robustly isolated from design choices.

free parameters (6)
  • t-SNE perplexity = 12
    Chosen by visual inspection of the scatterplot; no exhaustive search (Section 3.3.3).
  • K-means clusters for initial L0 = 20
    Chosen to balance diversity and feasibility; two images per cluster give 40 initial labeled images (Section 3.3.2).
  • Heatmap weighting exponent = 2
    Quadratic weighting (1 - avgConfidence)^2 was selected to emphasize uncertain images; Section 3.3.5.
  • Labeling budget per iteration = 30
    Chosen as a feasible annotation effort with expected impact on model performance (Section 3.2).
  • Fine-tuning epochs = 50
    Fixed based on preliminary testing and model stabilization (Section 3.3.2).
  • Image size and random seed = imgsz=640, seed=42
    Fixed to ensure reproducibility of the training regime (Sections 3.3.2, 3.6).
assumptions (4)
  • domain assumption t-SNE at perplexity 12 preserves enough local structure to guide useful sample selection.
    The exploration strategy and the data view rely on this; the paper notes the risk of distortion (Sections 1.6, 3.3.3).
  • domain assumption Feature embeddings extracted from the pretrained YOLOv11n backbone capture visual structure relevant to the active learning task.
    Sections 3.3.2 and 3.3.3 use these embeddings and t-SNE for all visualization and selection.
  • domain assumption The fixed 50-epoch YOLO fine-tuning protocol with seed 42 is a stable and unbiased training regime for all strategies.
    Section 3.3.2 fixes these parameters; no multiple runs are used to check variance.
  • domain assumption The author, acting as the sole expert user, executed the three predefined strategies faithfully and without bias.
    Sections 3.1 and 3.5 state the author is the single user, which is the core internal validity assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VILOD: A Visual Interactive Labeling Tool for Object Detection." pith.science (2026). https://pith.science/paper/ZMSBEDSF

@misc{pith2026250905317,
  author       = {Pith},
  title        = {Pith review of: VILOD: A Visual Interactive Labeling Tool for Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZMSBEDSF}},
  note         = {Machine review of arXiv:2509.05317}
}
read the original abstract

The advancement of Object Detection (OD) using Deep Learning (DL) is often hindered by the significant challenge of acquiring large, accurately labeled datasets, a process that is time-consuming and expensive. While techniques like Active Learning (AL) can reduce annotation effort by intelligently querying informative samples, they often lack transparency, limit the strategic insight of human experts, and may overlook informative samples not aligned with an employed query strategy. To mitigate these issues, Human-in-the-Loop (HITL) approaches integrating human intelligence and intuition throughout the machine learning life-cycle have gained traction. Leveraging Visual Analytics (VA), effective interfaces can be created to facilitate this human-AI collaboration. This thesis explores the intersection of these fields by developing and investigating "VILOD: A Visual Interactive Labeling tool for Object Detection". VILOD utilizes components such as a t-SNE projection of image features, together with uncertainty heatmaps and model state views. Enabling users to explore data, interpret model states, AL suggestions, and implement diverse sample selection strategies within an iterative HITL workflow for OD. An empirical investigation using comparative use cases demonstrated how VILOD, through its interactive visualizations, facilitates the implementation of distinct labeling strategies by making the model's state and dataset characteristics more interpretable (RQ1). The study showed that different visually-guided labeling strategies employed within VILOD result in competitive OD performance trajectories compared to an automated uncertainty sampling AL baseline (RQ2). This work contributes a novel tool and empirical insight into making the HITL-AL workflow for OD annotation more transparent, manageable, and potentially more effective.

Figures

Figures reproduced from arXiv: 2509.05317 by the authors.

Figure 1.1
Figure 1.1. Pool Based Active Learning Cycle. The effectiveness of the AL relies on the query strategy used to select the samples. Li et al. [16] defines five groups of common querying strategies: Uncertainty-Based Strategies. Uncertainty-based methods focus on selecting samples where the current deep learning model exhibits the highest ambiguity or lowest confidence in its predictions. Common techniques measure uncertainty thr… view at source ↗
Figure 2.2
Figure 2.2. VisGIL workspace by Grimmeisen et al. [27] [PITH_FULL_IMAGE:figures/full_fig_p020_2_2.png] view at source ↗
Figure 2.3
Figure 2.3. Overview of the dashboard by Klaassen [28] tive visual interface composed of multiple coordinated views, such as a pattern view based on UMAP and glyphs, and detailed instance views. A key contribution of SLAMVis is its novel pattern-based query strategy, which employs SOINN combined with K-means to identify data patterns and recommend informative candidate instances for labeling, aim￾ing to improve model effectiven… view at source ↗
Figures from the paper (70 more)
Figure 2.4
Figure 2.4. Figure 2.4: The Interactive Visual Interface of SLAMVis by Qian et al. [ [PITH_FULL_IMAGE:figures/full_fig_p021_2_4.png]
Figure 3.5
Figure 3.5. Figure 3.5: Overview of the suggested approach in the VILOD system. [PITH_FULL_IMAGE:figures/full_fig_p028_3_5.png]
Figure 3.6
Figure 3.6. Figure 3.6: Dashboard of Visual Interactive Labeling tool for Object Detection. (a) Dis [PITH_FULL_IMAGE:figures/full_fig_p031_3_6.png]
Figure 3.7
Figure 3.7. Figure 3.7: Training page of VILOD. (f) Displays the Model Training View, enabling real [PITH_FULL_IMAGE:figures/full_fig_p031_3_7.png]
Figure 3.8
Figure 3.8. Figure 3.8: The Scatterplot visualization in the Data View, initial state. [PITH_FULL_IMAGE:figures/full_fig_p032_3_8.png]
Figure 3.9
Figure 3.9. Figure 3.9: Workflow showing how the linkage between different views is utilized in the [PITH_FULL_IMAGE:figures/full_fig_p033_3_9.png]
Figure 3.10
Figure 3.10. Figure 3.10: The annotation interface of the VILOD system [PITH_FULL_IMAGE:figures/full_fig_p033_3_10.png]
Figure 3.11
Figure 3.11. Figure 3.11: The Labeled Images view and the removal of an image from the labeled pool. is a box plot visualizing the distribution of predictions of each class in the dataset. The purpose of the box plot is to give the user more granular insight into the model’s perfor￾mance on …
Figure 3.12
Figure 3.12. Figure 3.12: An overview of the System Architecture of VILOD [PITH_FULL_IMAGE:figures/full_fig_p035_3_12.png]
Figure 3.13
Figure 3.13. Figure 3.13: The Database Schema of the mySQL service. [PITH_FULL_IMAGE:figures/full_fig_p037_3_13.png]
Figure 4.14
Figure 4.14. Figure 4.14: Data View at the start of iteration 1 [PITH_FULL_IMAGE:figures/full_fig_p042_4_14.png]
Figure 4.15
Figure 4.15. Figure 4.15: Model View at the start of training iteration 1. [PITH_FULL_IMAGE:figures/full_fig_p042_4_15.png]
Figure 4.16
Figure 4.16. Figure 4.16: Three identified blobs within the Data View scatter plot. [PITH_FULL_IMAGE:figures/full_fig_p042_4_16.png]
Figure 4.17
Figure 4.17. Figure 4.17: First sample selection of iteration 1 [PITH_FULL_IMAGE:figures/full_fig_p043_4_17.png]
Figure 4.18
Figure 4.18. Figure 4.18: Annotating the first sample batch. Continuing the exploration of the Data View, another region of samples is found at the bottom of the fourth quadrant ( [PITH_FULL_IMAGE:figures/full_fig_p043_4_18.png]
Figure 4.19
Figure 4.19. Figure 4.19: Second sample selection and annotation. 39 [PITH_FULL_IMAGE:figures/full_fig_p043_4_19.png]
Figure 4.20
Figure 4.20. Figure 4.20: Class Balance after almost half the labeling budget is spent. [PITH_FULL_IMAGE:figures/full_fig_p044_4_20.png]
Figure 4.21
Figure 4.21. Figure 4.21: Exploring the Dataview, finding instances of rhinos and zebras. [PITH_FULL_IMAGE:figures/full_fig_p044_4_21.png]
Figure 4.22
Figure 4.22. Figure 4.22: Starting retraining and monitoring the training process. [PITH_FULL_IMAGE:figures/full_fig_p045_4_22.png]
Figure 4.23
Figure 4.23. Figure 4.23: Data View in the start of iteration 2 [PITH_FULL_IMAGE:figures/full_fig_p046_4_23.png]
Figure 4.24
Figure 4.24. Figure 4.24: Model View in the start of training iteration 2. [PITH_FULL_IMAGE:figures/full_fig_p046_4_24.png]
Figure 4.25
Figure 4.25. Figure 4.25: Data View of Iteration 2 with disabled heatmap. [PITH_FULL_IMAGE:figures/full_fig_p046_4_25.png]
Figure 4.26
Figure 4.26. Figure 4.26: A selection of a smaller region made in iteration 2. [PITH_FULL_IMAGE:figures/full_fig_p047_4_26.png]
Figure 4.27
Figure 4.27. Figure 4.27: Four selections made in iteration 2. Targeting areas with no previously [PITH_FULL_IMAGE:figures/full_fig_p047_4_27.png]
Figure 4.28
Figure 4.28. Figure 4.28: Data View in the start of iteration 3 [PITH_FULL_IMAGE:figures/full_fig_p048_4_28.png]
Figure 4.29
Figure 4.29. Figure 4.29: Model View at the start of training iteration 3. [PITH_FULL_IMAGE:figures/full_fig_p048_4_29.png]
Figure 4.30
Figure 4.30. Figure 4.30: Four selections made in iteration 2. Targeting areas with no previously [PITH_FULL_IMAGE:figures/full_fig_p049_4_30.png]
Figure 4.31
Figure 4.31. Figure 4.31: Data View at the start of iteration 4, without heatmap [PITH_FULL_IMAGE:figures/full_fig_p049_4_31.png]
Figure 4.32
Figure 4.32. Figure 4.32: A selection of three points closely coupled together but separated from the [PITH_FULL_IMAGE:figures/full_fig_p050_4_32.png]
Figure 4.33
Figure 4.33. Figure 4.33: Model View at the start of training iteration 5. [PITH_FULL_IMAGE:figures/full_fig_p050_4_33.png]
Figure 4.34
Figure 4.34. Figure 4.34: Two selections made in iteration 5. Targeting areas with no previously la [PITH_FULL_IMAGE:figures/full_fig_p051_4_34.png]
Figure 4.35
Figure 4.35. Figure 4.35: Isolated AL samples and Uncertainty heatmap and the first sample selection [PITH_FULL_IMAGE:figures/full_fig_p052_4_35.png]
Figure 4.36
Figure 4.36. Figure 4.36: Data View iteration 2 [PITH_FULL_IMAGE:figures/full_fig_p053_4_36.png]
Figure 4.37
Figure 4.37. Figure 4.37: Model View in the start of training iteration 2. [PITH_FULL_IMAGE:figures/full_fig_p053_4_37.png]
Figure 4.38
Figure 4.38. Figure 4.38: Example of a sample suggested by the Active Learning algorithm. [PITH_FULL_IMAGE:figures/full_fig_p054_4_38.png]
Figure 4.39
Figure 4.39. Figure 4.39: Data View iteration 3 A common feature of the AL samples this round is that they contain many instances of animals, often times also obscured or far in the background. One such example can be seen in [PITH_FULL_IMAGE:figures/full_fig_p054_4_39.png]
Figure 4.40
Figure 4.40. Figure 4.40: Model View at the start of training iteration 2. [PITH_FULL_IMAGE:figures/full_fig_p055_4_40.png]
Figure 4.41
Figure 4.41. Figure 4.41: Example of a noisy sample with many partially hidden instances of the Buffalo class [PITH_FULL_IMAGE:figures/full_fig_p055_4_41.png]
Figure 4.43
Figure 4.43. Figure 4.43: Data View iteration 4 [PITH_FULL_IMAGE:figures/full_fig_p056_4_43.png]
Figure 4.44
Figure 4.44. Figure 4.44: Model View at the start of training iteration 4. [PITH_FULL_IMAGE:figures/full_fig_p056_4_44.png]
Figure 4.45
Figure 4.45. Figure 4.45: A confidently misclassified sample. Although the Prediction Confidence Distribution in the Model View points to rhinos being the most confidently predicted class, the interplay between the different visualiza￾tions and inspecting AL suggestions tells a more nuanced …
Figure 4.46
Figure 4.46. Figure 4.46: Data View iteration 5 [PITH_FULL_IMAGE:figures/full_fig_p057_4_46.png]
Figure 4.47
Figure 4.47. Figure 4.47: Model View at the start of training iteration 5. [PITH_FULL_IMAGE:figures/full_fig_p057_4_47.png]
Figure 4.48
Figure 4.48. Figure 4.48: AL samples in iteration 5. Labeling this final round of AL samples and training the last iteration of the model with a total of 190 labeled images in the training set concludes this use case, which fo￾cused on uncertainty sampling as facilitated by the VILOD tool. T…
Figure 4.49
Figure 4.49. Figure 4.49: First selection of samples in iteration 1. [PITH_FULL_IMAGE:figures/full_fig_p059_4_49.png]
Figure 4.50
Figure 4.50. Figure 4.50: Selection of samples depicting rhinos from a side view perspective. [PITH_FULL_IMAGE:figures/full_fig_p059_4_50.png]
Figure 4.51
Figure 4.51. Figure 4.51: Class Balance after half the labeling budget is spent. [PITH_FULL_IMAGE:figures/full_fig_p060_4_51.png]
Figure 4.52
Figure 4.52. Figure 4.52: Final selection and annotation for iteration 1. [PITH_FULL_IMAGE:figures/full_fig_p060_4_52.png]
Figure 4.53
Figure 4.53. Figure 4.53: Data view in iteration 2 [PITH_FULL_IMAGE:figures/full_fig_p061_4_53.png]
Figure 4.54
Figure 4.54. Figure 4.54: Model View at the start of iteration 2. pool of water grouped together. Additionally, a few similar-looking samples of elephants are in some kind of enclosed space, with man-made structures in the background. This observation highlights the benefit of using a pre-tr…
Figure 4.55
Figure 4.55. Figure 4.55: Selection of a high uncertainty region in the first quadrant. [PITH_FULL_IMAGE:figures/full_fig_p061_4_55.png]
Figure 4.56
Figure 4.56. Figure 4.56: AL suggested sample with misclassified predictions made by the model. [PITH_FULL_IMAGE:figures/full_fig_p062_4_56.png]
Figure 4.57
Figure 4.57. Figure 4.57: Second selection and another misclassified sample. [PITH_FULL_IMAGE:figures/full_fig_p062_4_57.png]
Figure 4.58
Figure 4.58. Figure 4.58: Data View start of iteration 3 [PITH_FULL_IMAGE:figures/full_fig_p063_4_58.png]
Figure 4.59
Figure 4.59. Figure 4.59: Model View start of iteration 3. 59 [PITH_FULL_IMAGE:figures/full_fig_p063_4_59.png]
Figure 4.60
Figure 4.60. Figure 4.60: Selecting AL samples in a high uncertainty region. [PITH_FULL_IMAGE:figures/full_fig_p064_4_60.png]
Figure 4.62
Figure 4.62. Figure 4.62: A sample containing both buf￾falos and zebras 60 [PITH_FULL_IMAGE:figures/full_fig_p064_4_62.png]
Figure 4.63
Figure 4.63. Figure 4.63: A discarded AL sample with noisy features. [PITH_FULL_IMAGE:figures/full_fig_p065_4_63.png]
Figure 4.64
Figure 4.64. Figure 4.64: The Model Validation view after performing three rounds of iterative train [PITH_FULL_IMAGE:figures/full_fig_p065_4_64.png]
Figure 4.65
Figure 4.65. Figure 4.65: Data View start of iteration 4 [PITH_FULL_IMAGE:figures/full_fig_p066_4_65.png]
Figure 4.66
Figure 4.66. Figure 4.66: Model View start of iteration 4. 62 [PITH_FULL_IMAGE:figures/full_fig_p066_4_66.png]
Figure 4.67
Figure 4.67. Figure 4.67: A selection of points in the fourth iteration and the inspection of a sample of [PITH_FULL_IMAGE:figures/full_fig_p067_4_67.png]
Figure 4.68
Figure 4.68. Figure 4.68: A selection of points in the fourth iteration and the inspection and correction [PITH_FULL_IMAGE:figures/full_fig_p067_4_68.png]
Figure 4.69
Figure 4.69. Figure 4.69: Model Validation View at the start of iteration 5. [PITH_FULL_IMAGE:figures/full_fig_p068_4_69.png]
Figure 4.70
Figure 4.70. Figure 4.70: Data View start of iteration 5. 64 [PITH_FULL_IMAGE:figures/full_fig_p068_4_70.png]
Figure 4.71
Figure 4.71. Figure 4.71: Model View start of iteration 5. The sampling strategy for this iteration will be similar to that of iteration 4. An attempt will be made to find underrepresented regions, but AL samples will also be inspected, and those deemed to indicate a structural confusion of …
Figure 4.72
Figure 4.72. Figure 4.72: Selection of a cluster separated from other points. [PITH_FULL_IMAGE:figures/full_fig_p069_4_72.png]
Figure 4.74
Figure 4.74. Figure 4.74: AL sample with noisy detections stacked upon each other [PITH_FULL_IMAGE:figures/full_fig_p070_4_74.png]
Figure 4.75
Figure 4.75. Figure 4.75: Four examples of samples with multiple instances of the elephant class and [PITH_FULL_IMAGE:figures/full_fig_p070_4_75.png]
Figure 4.76
Figure 4.76. Figure 4.76: mAP50-95 performance trajectory comparison between the different label [PITH_FULL_IMAGE:figures/full_fig_p072_4_76.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages

  1. [1]

    Object detection in 20 years: A survey,

    Z. Zou, K. Chen, Z. Shi, Y . Guo, and J. Ye, “Object detection in 20 years: A survey,” Proceedings of the IEEE, vol. 111, no. 3, pp. 257–276, 2023

    work page 2023

  2. [2]

    A comprehensive survey on deep active learning in medical image analysis,

    H. Wang, Q. Jin, S. Li, S. Liu, M. Wang, and Z. Song, “A comprehensive survey on deep active learning in medical image analysis,”Medical Image Analysis, p. 103201, 2024

    work page 2024

  3. [3]

    Active learning literature survey,

    B. Settles, “Active learning literature survey,” 2009

    2009

  4. [4]

    A survey of human-in-the-loop for machine learning,

    X. Wu, L. Xiao, Y . Sun, J. Zhang, T. Ma, and L. He, “A survey of human-in-the-loop for machine learning,”Future Generation Computer Systems, vol. 135, pp. 364–381, 2022

    2022

  5. [5]

    J. J. Thomas and K. A. Cook, Eds., Illuminating the Path: The Research and Devel- opment Agenda for Visual Analytics. IEEE Computer Society Press, 2005

    work page 2005

  6. [6]

    A trajectory scoring tool for local anomaly detection in maritime traffic using visual analytics,

    F. H. Abreu, A. Soares, F. V . Paulovich, and S. Matwin, “A trajectory scoring tool for local anomaly detection in maritime traffic using visual analytics,” ISPRS Inter- national Journal of Geo-Information, vol. 10, no. 6, p. 412, 2021

    work page 2021

  7. [7]

    Power to the people: The role of humans in interactive machine learning,

    S. Amershi, M. Cakmak, W. B. Knox, and T. Kulesza, “Power to the people: The role of humans in interactive machine learning,” AI magazine, vol. 35, no. 4, pp. 105–120, 2014

    work page 2014

  8. [8]

    A comprehensive survey of machine learning techniques and models for object detection,

    M. Trigka and E. Dritsas, “A comprehensive survey of machine learning techniques and models for object detection,” Sensors, vol. 25, no. 1, 2025. [Online]. Available: https://www.mdpi.com/1424-8220/25/1/214

    work page 2025

Show all 34 references
  1. [9]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” nature, vol. 521, no. 7553, pp. 436–444, 2015

    2015

  2. [10]

    A survey of deep learning-based object detection,

    L. Jiao, F. Zhang, F. Liu, S. Yang, L. Li, Z. Feng, and R. Qu, “A survey of deep learning-based object detection,” IEEE Access, vol. 7, pp. 128 837–128 868, 2019

    2019

  3. [11]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779–788

    2016

  4. [12]

    Ssd: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer, 2016, pp. 21–37

    2016

  5. [13]

    Ultralytics yolo11,

    G. Jocher and J. Qiu, “Ultralytics yolo11,” 2024. [Online]. Available: https: //github.com/ultralytics/ultralytics

    2024

  6. [14]

    YOLO performance metrics

    Ultralytics. YOLO performance metrics. Ultralytics. Accessed: 2025-05-07. [Online]. Available: https://docs.ultralytics.com/guides/yolo-performance-metrics/

    2025

  7. [15]

    Vista: A visual ana- lytics platform for semantic annotation of trajectories,

    A. Soares, J. Rose, M. Etemad, C. Renso, and S. Matwin, “Vista: A visual ana- lytics platform for semantic annotation of trajectories,” in Proceedings of the 22nd International Conference on Extending Database Technology (EDBT), 2019. 76

    2019

  8. [16]

    A survey on deep active learning: Recent advances and new frontiers,

    D. Li, Z. Wang, Y . Chen, R. Jiang, W. Ding, and M. Okumura, “A survey on deep active learning: Recent advances and new frontiers,” IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  9. [17]

    Analytic: An active learning system for trajectory classification,

    A. S. Júnior, C. Renso, and S. Matwin, “Analytic: An active learning system for trajectory classification,” IEEE computer graphics and applications, vol. 37, no. 5, pp. 28–39, 2017

    2017

  10. [18]

    Local anomaly detection in maritime traffic using visual analytics

    F. H. O. Abreu, A. Soares, F. V . Paulovich, and S. Matwin, “Local anomaly detection in maritime traffic using visual analytics.” inIn: EDBT/ICDT Workshops: 4th Inter- national Workshop on Big Mobility Data Analytics (BMDA 2021), Nicosia, Cyprus., 2021

    2021

  11. [19]

    A review of user interface design for interac- tive machine learning,

    J. J. Dudley and P. O. Kristensson, “A review of user interface design for interac- tive machine learning,” ACM Transactions on Interactive Intelligent Systems (TiiS), vol. 8, no. 2, pp. 1–37, 2018

    2018

  12. [20]

    Rajaraman and J

    A. Rajaraman and J. D. Ullman, Mining of massive datasets. Autoedicion, 2011

    2011

  13. [21]

    Visualizing data using t-sne,

    L. v. d. Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of machine learning research, vol. 9, no. Nov, pp. 2579–2605, 2008

    2008

  14. [22]

    Umap: Uniform manifold approximation and projection for dimension reduction,

    L. McInnes, J. Healy, and J. Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,” arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  15. [23]

    A compre- hensive survey on transfer learning,

    F. Zhuang, Z. Qi, K. Duan, D. Xi, Y . Zhu, H. Zhu, H. Xiong, and Q. He, “A compre- hensive survey on transfer learning,” Proceedings of the IEEE, vol. 109, no. 1, pp. 43–76, 2020

    2020

  16. [24]

    Examining the impact of cross-domain learning on crime prediction,

    F. K. Bappee, A. Soares, L. M. Petry, and S. Matwin, “Examining the impact of cross-domain learning on crime prediction,” Journal of big data, vol. 8, no. 1, p. 96, 2021

    2021

  17. [25]

    Visually-enabled active deep learning for (geo) text and image classification: a review,

    L. Yang, A. M. MacEachren, P. Mitra, and T. Onorati, “Visually-enabled active deep learning for (geo) text and image classification: a review,”ISPRS International Journal of Geo-Information, vol. 7, no. 2, p. 65, 2018

    2018

  18. [26]

    Comparing visual-interactive labeling with active learning: An experimental study,

    J. Bernard, M. Hutter, M. Zeppelzauer, D. Fellner, and M. Sedlmair, “Comparing visual-interactive labeling with active learning: An experimental study,”IEEE trans- actions on visualization and computer graphics, vol. 24, no. 1, pp. 298–308, 2017

    2017

  19. [27]

    Visgil: machine learning-based vi- sual guidance for interactive labeling,

    B. Grimmeisen, M. Chegini, and A. Theissler, “Visgil: machine learning-based vi- sual guidance for interactive labeling,” The Visual Computer, vol. 39, no. 10, pp. 5097–5119, 2023

    2023

  20. [28]

    Interactive visualization for multi-target active learning,

    E. T. C. Klaassen, “Interactive visualization for multi-target active learning,” Master Thesis, Eindhoven University of Technology, Eindhoven, The Netherlands, 2024, department of Mathematics and Computer Science, Visualization Group

    2024

  21. [29]

    Slamvis: An interactive visualization approach for smart labeling on multidimensional data,

    A. Qian, C. Li, X. Dong, S. Chen, Y . Zhang, and G. Li, “Slamvis: An interactive visualization approach for smart labeling on multidimensional data,” in 2021 IEEE Intl Conf on Parallel & Distributed Processing with Applications, Big Data & Cloud Computing, Sustainable Computin...

    2021

  22. [30]

    Active learning for interactive visualiza- tion,

    T. Iwata, N. Houlsby, and Z. Ghahramani, “Active learning for interactive visualiza- tion,” in Artificial Intelligence and Statistics. PMLR, 2013, pp. 342–350

    2013

  23. [31]

    Active learning for deep object detection,

    C.-A. Brust, C. Käding, and J. Denzler, “Active learning for deep object detection,”

  24. [32]

    Semi-supervised active learning for object detec- tion,

    S. Chen, Y . Yang, and Y . Hua, “Semi-supervised active learning for object detec- tion,” Electronics, vol. 12, no. 2, p. 375, 2023

    2023

  25. [33]

    Kernel density estimation — scipy v1.14.0 manual,

    SciPy Developers, “Kernel density estimation — scipy v1.14.0 manual,” https:// docs.scipy.org/doc/scipy/tutorial/stats/kernel_density_estimation.html, 2025. 78 A Appendix 1 Use Case 1: Exploration & Structure Focus Annotated Samples Iteration 1 Table 1.4: Exploration & Structu...

    2025

  26. [2018]

    Available: https://arxiv.org/abs/1809.09875

    [Online]. Available: https://arxiv.org/abs/1809.09875

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.