Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Diffuse the object, keep its label: curating detector training data from a few unlabeled photographs via VLM-built 3D vegetation scenes

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A detector trained on synthetic images made from a few unlabeled site photographs matched or exceeded a detector trained on a much larger real labeled dataset, for cross-site landmine detection.

desk verdict Graded mask-lock for diffusion-refined synthetic data is a real new idea, but the headline SULAND margin is the best of four alpha values chosen on the test split—read it as a promising pilot, not a fixed-protocol result. read the letter →

arxiv 2608.09691 v1 pith:R3LGYVRC submitted 2026-08-10 cs.CV

Mario Malizia , Marnix Enting , Rob Haelterman , Ken Hasselmann This is my paper · ORCID
classification cs.CV
keywords synthetictrainingdatadiffusionmodelshumanitariandeminingobjectdetectioncross-sitegeneralizationdomainadaptation3Dscenesynthesismask-lockedcompositing
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

The paper is trying to establish that for detecting small or camouflaged objects hidden in vegetation, a detector can be trained on synthetic images generated from a handful of unlabeled photographs of the deployment site itself, and that this training set can generalize across sites as well as or better than a far larger set of labeled real images from a different site. If true, the per-site data question for such detection pipelines would shift from collecting and labeling new frames to photographing the site and regenerating the training set. The central mechanism is a graded mask-lock parameter that determines how much of the rendered object a diffusion model may re-texture, so labels remain valid by construction while the object's appearance is transferred. The paper is careful to present the finding as preliminary evidence on one benchmark, one object family, and one detector, not as a general claim.

What carries the argument

The central object is the graded mask-lock compositing step, a single parameter $\alpha \in [0,1]$. The pipeline first uses a vision-language model to turn one reference photograph into a coarse Blender scene of vegetation primitives, places 3D meshes of the target objects so that boxes, masks, and per-instance occlusion follow from scene geometry, and then passes each render through an SDEdit diffusion pass (an image-editing method that re-noises and denoises a render in latent space) conditioned on a terrain LoRA fine-tuned from unlabeled crops. The mask-lock then pastes the rendered object pixels back through the visible-surface mask: at $\alpha=0$ the object is fully protected from diffusion, at $\alpha=1$ diffusion touches everything, and intermediate values feather the mask so a controlled share of the object's appearance comes from the diffusion pass while its geometry and label stay fixed. This mechanism is what keeps labels valid by construction while still transferring the surface appearance that low-contrast objects need for detection.

What would settle it

Deliberately perturb the estimated camera height or object scale in the scene-generation step by about 30 percent, keep every other setting identical, and measure cross-site macro-F1 on SULAND USA across five seeds; if the score does not drop outside seed variance, the approximate-camera assumption is not load-bearing, and if it does, the geometric scaffold is the reason the pipeline works.

Watch

Extended reading notes

Core claim

The paper claims that on the SULAND benchmark, training a YOLOv11-L detector on 1,000 synthetic images produced by this pipeline generalizes across sites at least as well as training the same detector on the 22,756 labeled real frames of the ITA split: the reference configuration reaches macro-F1 $0.584 \pm 0.025$ on the USA test split, versus $0.393 \pm 0.021$ for the real labeled baseline, with the difference several times the seed standard deviation. The paper reads this as preliminary evidence, not a general claim about synthetic data superiority, and emphasizes that the comparison is unsupervised adaptation to the target site from a handful of unlabeled photographs versus conventional cross-site label reuse. It further claims that the mask-lock grade is the most influential curation choice: lightly diffusing the object at $\alpha=0.10$ roughly doubles PMA-2 recall relative to full pixel protection, while unrestricted diffusion dissolves the object and gives back most of the gain.

Load-bearing premise

The whole approach rests on the assumption that a coarse 3D scene built by a vision-language model from one photograph, with the camera distance only roughly known, is a faithful enough scaffold that the diffusion re-texturing transfers the site's appearance while the mask-lock keeps the object's label valid.

Editorial extensions

If this is right

  • The mask-lock sweep shows an inverted-U: mF1 rises from 0.441 at $\alpha=0$ to 0.584 at $\alpha=0.10$ and falls to 0.466 at $\alpha=1$, so the curation dial itself shifts generalization more than any data-budget choice tested.
  • The budget axis is flat within error: a single reference photograph reaches $0.562 \pm 0.011$ mF1 at the reference lock, within one standard deviation of the ten-photograph configuration, so per-site data collection can plausibly drop to one photo for approximately uniform terrain.
  • In-domain validation mAP50 is at least 0.99 for nearly all synthetic configurations while real-domain mF1 spans 0.05 to 0.62, so in-domain accuracy cannot indicate when a model is ready to generalize; downstream evaluation on the real target site remains the only reliable guide.
  • Raw CG renders without diffusion reach only $0.194 \pm 0.111$ mF1, so the diffusion re-texturing and the mask-lock, not the 3D rendering alone, carry the cross-site generalization gain.

Reading between the lines

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

  • The same recipe should transfer to other detection settings with a small, geometrically defined object catalog and scarce real data, such as invasive-species monitoring or crop and weed detection, provided a terrain LoRA and a mesh library can be assembled; the paper states the pipeline is not tied to landmines but does not evaluate this.
  • Because one photograph matched ten photographs within error, the practical bottleneck is likely object-mesh fidelity and scene-geometry accuracy rather than data-collection budget, which points toward range-sensing support for the scene scaffold as a robustness upgrade, a direction the paper lists only as future work.
  • A natural next comparison would pit this pipeline against established unsupervised domain-adaptation baselines such as style transfer, copy-paste synthesis, or Gaussian-splatting insertion on the same SULAND split; the paper does not run that comparison, so its claim is specifically about beating label reuse, not about beating all adaptation methods.
  • The paper's probe metrics (Kernel Inception Distance and open-vocabulary recall) did not predict cross-site mF1, suggesting that designing a cheap, calibrated generalization predictor is an open problem; a systematic sweep over mask-lock grades and terrain pairs could reveal whether the optimal grade is site-dependent.
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 / 5 minor

Summary. The paper proposes a pipeline for generating labeled detector training images from a handful of unlabeled photographs of a deployment site. A VLM produces a coarse Blender scene from one reference photo; scripted placement of 3D ordnance meshes yields boxes, masks, and per-instance occlusion; a LoRA fine-tuned on unlabeled crops conditions an SDEdit pass that re-textures the renders; and a graded mask-lock parameter alpha controls how much diffusion may touch the object. The authors train YOLOv11-L on the resulting synthetic pool and evaluate cross-site generalization on the SULAND benchmark (ITA to USA). The headline result is that the reference synthetic configuration (M2, alpha=0.10) reaches mF1=0.584 +/- 0.025 versus 0.393 +/- 0.021 for a detector trained on the real labeled ITA split. Budget sweeps suggest that the mask-lock grade matters more than photograph/crop/image-count budgets, and in-domain validation mAP is shown to be uninformative about cross-site performance.

Significance. If the central claim survives scrutiny, the pipeline is a useful contribution to data-scarce detection in vegetation, particularly humanitarian demining: it produces exact labels and occlusion from 3D geometry, requires only a handful of unlabeled site photographs, and the graded mask-lock idea is a sensible answer to the failure mode of unconstrained diffusion destroying the target. The paper has genuine methodological strengths: five training seeds per configuration, a shared render pool that removes scene variance across ablation cells, explicit disclosure that the synthetic pipeline consumes test-site unlabeled photographs, a matched-frame exclusion check, and an unusually candid limitations section. The authors also report negative results (KID and Grounding-DINO probes do not predict generalization), which is honest and informative. However, the headline comparison is weakened by selecting the mask-lock grade on the same test split used for the final claim, and the abstract's phrasing overstates the comparison by not consistently foregrounding that the synthetic pipeline uses target-site unlabeled data.

major comments (3)
  1. [§4.3, Table 1] The reference configuration M2 (alpha=0.10) is not pre-specified; it was selected as the best of four mask-lock grades evaluated on the same USA test split used for the headline comparison. Section 4.4 states that "within this study, curation decisions had to be guided by downstream evaluation," confirming that alpha was chosen by test-domain mF1. The reported +/-0.025 seed standard deviation is within-configuration variance and does not account for selection over four alpha values. Since M0, M1, and M3 score 0.441, 0.503, and 0.466, respectively, only M2 clearly exceeds R1 (0.393); the claim that the pipeline "matched or exceeded" the real baseline therefore rests on the best of four settings. Please select alpha without access to the test split (e.g., on a separate validation site, by pre-registration, or by demonstrating robustness across a plausible alpha interval) and report a selection-adjusted or sensitivity analysis.
  2. [§4.2 and Appendix A.4] The comparison is not purely "synthetic data versus real labeled data": the synthetic pipeline consumes ten unlabeled photographs from the USA evaluation site, one as the Stage 1 scene reference and all for LoRA fine-tuning and sensor-noise calibration. The paper discloses this and frames the result as unsupervised site adaptation, but the abstract and Section 6 still say that a detector trained on "our synthetic data" matched or exceeded a detector trained on real labeled data from a different site. Please make the headline claim consistently state that the comparison is unsupervised adaptation with target-site unlabeled photographs versus conventional cross-site label reuse; otherwise readers will over-interpret the result as evidence of synthetic-data superiority.
  3. [§4.3, Table 2] The claim that gains are "largely insensitive to the photograph and crop budgets" is only partially supported. The full photograph and patch sweeps are run at alpha=0, and only the extreme cells were re-measured at alpha=0.10; P2, C2, S2, and S3 are marked n/a at the reference lock. The abstract's budget-insensitivity statement therefore holds only for the hard-lock setting plus two re-measured extremes. Please either complete the alpha=0.10 sweep for the missing cells or soften the claim to the settings actually measured.
minor comments (5)
  1. [Table 1, footnote] The tie-band footnote defines Delta<0.01 on per-class recall, but per-class recall values are reported without seed variance; please clarify how the tie band is computed and whether it is applied to point estimates or to some measure of uncertainty.
  2. [Appendix A.1] The sun-strength expression "e0 * 2.5 * (0.6 + 1.4 * L_bar)" is not fully defined; please define e0 and the units/range of L_bar (mean luminance) so that the capture-realism calibration is reproducible.
  3. [Appendix A.4] The sentence "Removing the matched frames from the evaluation set is a null effect on the model the concern applies to" is convoluted; please rephrase to state plainly that the matched-frame exclusion was tested on the M2 configuration and left both mF1 and mAP50 unchanged.
  4. [Section 4.4] The statement that "real-domain mF1 spanned 0.05 to 0.62" is not directly visible in Tables 1 and 2, whose reported means start at 0.194; please clarify whether 0.05 refers to a per-seed minimum or to a different configuration not shown in the tables.
  5. [General] Consider adding a reproducibility statement with code and configuration release details; the appendix is detailed, but no link or artifact is provided.

Circularity Check

1 steps flagged · score 5.0 of 10

Headline advantage rests on selecting the mask-lock grade on the USA test split, making the reference configuration a test-fitted rather than fixed-protocol result.

  1. fitted input called prediction [Section 4.2 (Table 1) and Section 4.4; alpha selection on USA test split]
    "The reference synthetic configuration (M2, α=0.10) reaches mF1= 0.584±0.025, against 0.393±0.021 for the real labeled baseline (R1). [...] Within this study, curation decisions had to be guided by downstream evaluation."

    Table 1 evaluates four mask-lock grades (M0–M3, α=0, 0.05, 0.10, 1) on the same USA split that defines the headline result; M2 is then designated 'reference' after it yields the best mF1 (0.584 vs 0.441/0.503/0.466). The paper's own statement that 'curation decisions had to be guided by downstream evaluation' admits that α was selected using test-domain mF1. The central claim of matching or exceeding the real baseline is therefore a best-of-four test-set selection, not an out-of-sample prediction of a pre-specified pipeline. The reported ±0.025 is within-configuration seed variance and does not account for selection over four α values; only the chosen α clearly clears the real baseline (0.393).

full rationale

The pipeline itself is not definitionally circular: the VLM scene generation, LoRA adaptation, SDEdit re-texturing, and mask-locked compositing are all implemented as described, and the labels follow from 3D geometry rather than from the detector or the benchmark. There is no load-bearing self-citation chain and no uniqueness theorem imported from the authors' prior work. However, the paper's central quantitative claim is weakened by a genuine selection-on-test-set circularity: the reference mask-lock grade α=0.10 was chosen after evaluating four grades on the USA evaluation split, and the same split then supports the headline comparison against the real labeled baseline. The paper discloses this ('curation decisions had to be guided by downstream evaluation'), but disclosure does not remove the problem that the reported advantage is a maximum over test-evaluated configurations. A fair fixed-protocol comparison would require α to be pre-registered, selected on a separate validation split, or shown to be robust across the sampled range without peaking at the test-optimal value. Because the central 'matched or exceeded' claim is partly constituted by this test-set selection, the overall circularity score is moderate rather than zero; the underlying engineering is otherwise self-contained.

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

The central claim rests on a small number of domain assumptions and one fitted curation parameter. No invented entities are postulated.

free parameters (4)
  • Mask-lock grade alpha = 0.10 (reference)
    Swept over {0, 0.05, 0.10, 1} and the best value on the USA evaluation split was chosen as reference; the paper notes downstream evaluation guided curation.
  • Sensor-noise injection sigma = 15.8 (0-255 scale)
    Closed-loop calibration to match the measured target statistic sigma=12.3 on the reference photographs.
  • SDEdit strength = 0.42
    Fixed hand-set hyperparameter controlling how much the diffusion pass alters the render; not ablated.
  • Sun-strength scaling constants = e0 * 2.5 * (0.6 + 1.4 * L_bar)
    Formula calibrated to each reference photograph's mean luminance.
assumptions (4)
  • domain assumption The field's appearance is approximately stationary across the site, with near-constant vegetation, so a handful of unlabeled photographs capture the visual distribution.
    Stated in Section 3 Stage 2 as a working hypothesis that the field's appearance stays approximately stationary across the site, and in Section 5 limitation 3.
  • domain assumption The VLM's coarse Blender scene from a single photograph is a sufficient geometric scaffold for training a detector, including approximately correct camera pose and object scale.
    Section 3 Stage 1 samples the camera near the initial camera pose estimated by the VLM; Section 5 limitation 2 admits camera distance to ground is only approximately recovered while object scale matters for the graded lock.
  • domain assumption The label remains valid after mask-locked compositing because the Blender visible-surface mask is exact.
    Section 3 Stage 3 relies on the visible-surface mask exported by Blender to paste back object pixels; this treats the renderer's masks as ground truth.
  • domain assumption The LoRA trained on unlabeled crops captures terrain surface statistics without overfitting individual images.
    Section 3 Stage 2 assumes the small adapter captures the terrain's surface statistics rather than individual images, but no evidence rules out overfitting to the few crops.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffuse the object, keep its label: curating detector training data from a few unlabeled photographs via VLM-built 3D vegetation scenes." pith.science (2026). https://pith.science/paper/R3LGYVRC

@misc{pith2026260809691,
  author       = {Pith},
  title        = {Pith review of: Diffuse the object, keep its label: curating detector training data from a few unlabeled photographs via VLM-built 3D vegetation scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R3LGYVRC}},
  note         = {Machine review of arXiv:2608.09691}
}
read the original abstract

Labeled images of small objects hidden in vegetation are scarce, and detectors trained on them generalize poorly across sites. Rather than reusing labels collected at another site, we synthesize labeled training images from a handful of unlabeled photographs of the deployment site itself. A vision--language model generates a coarse 3D vegetation scene from one photograph; placing 3D object meshes in the scene yields bounding boxes, segmentation masks, and per-instance occlusion directly from the scene geometry, without manual annotation. A lightweight adapter fine-tuned on the photographs conditions a diffusion pass that re-textures the renders, and a graded mask-lock sets how much diffusion may touch the object itself. In our runs this grade was the most influential curation choice: lightly diffusing the object improves minority-class recall over fully protecting its pixels, while unrestricted diffusion dissolves it. Trained on these images, a standard detector matched or exceeded its counterpart trained on a larger labeled dataset of real images from a different site, consistently across seeds on a humanitarian-demining benchmark; the comparison is thus unsupervised site adaptation from a handful of photographs against conventional cross-site label reuse. In our ablations the gains were largely insensitive to the photograph and crop budgets, and in-domain accuracy did not predict cross-site performance.

Figures

Figures reproduced from arXiv: 2608.09691 by the authors.

Figure 1
Figure 1. Overview of the proposed pipeline. The real-data budget is a handful of un￾labeled photographs of the target site. One of them serves as the scene reference: a vision–language model emits a Blender Python script that approximates the terrain geometry from primitives (scene synthesis & rendering). 3D meshes of the target ordnance are placed in the resulting scene, so bounding boxes, instance masks, and per-instance v… view at source ↗
Figure 2
Figure 2. Mask-lock grade α across six scenes (rows), same diffusion seed per row. Columns: 3D render, occlusion-aware segmentation mask, lock grades in increasing order (α = 0, 0.05, 0.10, 1), and the reference output with its 2D labels (red: PFM-1, green: PMA-2). Rows 1–2: heavily occluded PFM-1s (41% and 51% visible); rows 3– 4: PMA-2 close-ups; rows 5–6: multi-landmine scenes. Unlocked, objects are absorbed into the backg… view at source ↗
Figure 3
Figure 3. The same comparison on the MineInsight forest-floor terrain (RGB stream; five scenes; leaf-litter LoRA), same columns as [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 36 canonical work pages

  1. [1]

    AI-Box: YOLOv5 – optimization of training data for PFM-1 antipersonnel mine detection.https://ai-box.eu/en/ai-pipeline-en/yolov5-optimization-of- training- data- for- pfm- 1- antipersonnel- mine- detection/872/(2023), ac- cessed 2026-07-10

    work page 2023

  2. [2]

    Remote Sensing16(12), 2046 (2024)

    Baur, J., Dewey, K., Steinberg, G., Nitsche, F.O.: Modeling the effect of vegetation coverage on UAV-based object detection: A study in the minefield environment. Remote Sensing16(12), 2046 (2024)

    work page 2024

  3. [3]

    In: ICLR (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD GANs. In: ICLR (2018)

    2018

  4. [4]

    Blender Online Community: Blender: a 3D modeling and rendering package.http: //www.blender.org(2024)

    work page 2024

  5. [5]

    Leafy Spurge Dataset: Real-world Weed Classification Within Aerial Drone Imagery

    Doherty,K.,Gurinas,M.,Samsoe,E.,Casper,C.,Larkin,B.,Ramsey,P.,Trabucco, B., Salakhutdinov, R.: Leafy spurge dataset: Real-world weed classification within aerial drone imagery. arXiv preprint arXiv:2405.03702 (2024)

    work page Pith review arXiv 2024

  6. [6]

    In: CVPR (2021)

    Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: CVPR (2021)

    work page 2021

  7. [7]

    Smart Agricul- tural Technology5, 100258 (2023)

    Giakoumoglou, N., Pechlivani, E.M., Tzovaras, D.: Generate-paste-blend-detect: Synthetic dataset for object detection in the agriculture domain. Smart Agricul- tural Technology5, 100258 (2023)

    work page 2023

  8. [8]

    Coarse-to-Real: Generative Rendering for Populated Dynamic Scenes

    Gomez-Nogales, G., Hong, Y., Ge, C., Zhuang, P., Comino-Trinidad, M., Casas, D., Zhou, Y.: Coarse-to-real: Generative rendering for populated dynamic scenes. arXiv preprint arXiv:2601.22301 (2026)

    work page Pith review arXiv 2026

Show all 42 references
  1. [9]

    In: IEEE ICRA Workshop on Field Robotics (2024)

    Hasselmann, K., Malizia, M., Caballero, R., Polisano, F., Govindaraj, S., Stigler, J., Ilchenko, O., Bajic, M., De Cubber, G.: A multi-robot system for the detection of explosive devices. In: IEEE ICRA Workshop on Field Robotics (2024)

    2024

  2. [10]

    arXiv preprint arXiv:2606.02580 (2026)

    He, G., Luo, R., Ma, W.C., Averbuch-Elor, H.: Thinking in Blender: Staged executable inverse graphics with vision-language models. arXiv preprint arXiv:2606.02580 (2026)

    2026 arXiv

  3. [11]

    In: ICLR (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)

    2022

  4. [12]

    In: ICML (2024)

    Hu, Z., Iscen, A., Jain, A., Kipf, T., Yue, Y., Ross, D.A., Schmid, C., Fathi, A.: SceneCraft: An LLM agent for synthesizing 3D scenes as Blender code. In: ICML (2024)

    2024

  5. [13]

    Computer Vision and Image Under- standing64(2), 300–302 (1996)

    Immerkær, J.: Fast noise variance estimation. Computer Vision and Image Under- standing64(2), 300–302 (1996)

    1996

  6. [14]

    https://www.mineactionstandards.org/standards/08-10/(2019), UN Mine Ac- tion Service (UNMAS)

    International Mine Action Standards (IMAS): IMAS 08.10: Non-technical survey. https://www.mineactionstandards.org/standards/08-10/(2019), UN Mine Ac- tion Service (UNMAS)

    2019

  7. [15]

    arXiv preprint arXiv:2512.08747 (2025)

    Károly, A.I., Galambos, P.: A scalable pipeline combining procedural 3D graphics and guided diffusion for photorealistic synthetic training data generation in white button mushroom segmentation. arXiv preprint arXiv:2512.08747 (2025)

    2025

  8. [16]

    In: ICLR (2026) 16 M

    Kim, D., Lee, J., Choi, J., Park, J., Lee, G., Kim, Y., Ki, M., Lee, J., Cha, J.: ForestPersons: A large-scale dataset for under-canopy missing person detection. In: ICLR (2026) 16 M. Malizia et al

    2026

  9. [17]

    Master’s thesis, Politecnico di Milano (2025), https://www.politesi.polimi.it/handle/10589/235105

    Lain,L.:GeneratingSyntheticDataforIllegalLandfillDetection:ADualApproach Using Inpainting and Cut-and-Paste. Master’s thesis, Politecnico di Milano (2025), https://www.politesi.polimi.it/handle/10589/235105

    2025

  10. [18]

    Agriculture15(2), 138 (2025)

    Li, T., Burridge, J., Blok, P.M., Guo, W.: A patch-level data synthesis pipeline enhances species-level crop and weed segmentation in natural agricultural scenes. Agriculture15(2), 138 (2025)

    2025

  11. [19]

    In: ECCV

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV. pp. 740–755 (2014)

    2014

  12. [20]

    In: ECCV (2024)

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., Zhu, J., Zhang, L.: Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In: ECCV (2024)

    2024

  13. [21]

    In: CVPR Work- shops (2026)

    Ljungbergh, W., Taveira, B., Zheng, W., Tonderski, A., Peng, C., Kahl, F., Peters- son, C., Felsberg, M., Keutzer, K., Tomizuka, M., Zhan, W.: R3D2: Realistic 3D asset insertion via diffusion for autonomous driving simulation. In: CVPR Work- shops (2026)

    2026

  14. [22]

    In: ICCV

    Lu, J., Xie, R., Xie, Z., Wu, Z., Xi, D., Ye, Q., Wang, R., Bao, H., Huo, Y.: IntrinsicControlNet: Cross-distribution image generation with real and unreal. In: ICCV. pp. 27315–27325 (2025)

    2025

  15. [23]

    IEEE Robotics and Automation Letters11(2), 1650–1657 (2026)

    Malizia, M., Hamesse, C., Hasselmann,K., De Cubber,G., Tsiogkas, N., Demeester, E.,Haelterman,R.:MineInsight:Amulti-sensordatasetforhumanitariandemining robotics in off-road environments. IEEE Robotics and Automation Letters11(2), 1650–1657 (2026)

    2026

  16. [24]

    Malizia,M.,Hasselmann,K.,Miuccio,A.,Haelterman,R.,Tsiogkas,N.,Demeester, E.: PFM-1 landmine detection in vegetation using thermal imaging with limited training data. In: Int. Conf. Control, Automation and Systems (ICCAS). pp. 1864– 1869 (2025)

    2025

  17. [25]

    In: ICLR (2022)

    Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)

    2022

  18. [26]

    In: Architecture of Computing Systems (ARCS)

    Modak, S., Saltık, A.O., Stein, A.: Exploring model quantization in GenAI-based image inpainting and detection of arable plants. In: Architecture of Computing Systems (ARCS). LNCS, Springer (2025)

    2025

  19. [27]

    arXiv preprint arXiv:2412.06248 (2024)

    Patwari, K., Schneider, D., Sun, X., Chuah, C.N., Lyu, L., Sharma, V.: Rendering- refined Stable Diffusion for privacy compliant synthetic data. arXiv preprint arXiv:2412.06248 (2024)

    2024 arXiv

  20. [28]

    ACM Transactions on Graphics22(3), 313–318 (2003)

    Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. ACM Transactions on Graphics22(3), 313–318 (2003)

    2003

  21. [29]

    In: ICLR (2024)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution im- age synthesis. In: ICLR (2024)

    2024

  22. [30]

    In: IEEE Int

    Prakash, A., Boochoon, S., Brophy, M., Acuna, D., Cameracci, E., State, G., Shapira, O., Birchfield, S.: Structured domain randomization: Bridging the reality gap by context-aware synthetic data. In: IEEE Int. Conf. Robotics and Automation (ICRA) (2019)

    2019

  23. [31]

    Roboflow Universe: Synthetic PFM-1 dataset.https://universe.roboflow.com/ wandering/synthetic-pfm-1-dataset-bwvyo/dataset/1(2025), accessed 2026- 07-10

    2025

  24. [32]

    In: CVPR (2022) Diffuse the Object, Keep Its Label 17

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) Diffuse the Object, Keep Its Label 17

    2022

  25. [33]

    ISPRS In- ternational Journal of Geo-Information13(7), 259 (2024)

    Saliba, A., Tout, K., Zaki, C., Claramunt, C.: Bridging human expertise with machine learning and GIS for mine type prediction and classification. ISPRS In- ternational Journal of Geo-Information13(7), 259 (2024)

    2024

  26. [34]

    In: CVPR (2023)

    Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: ObjectStitch: Object compositing with diffusion model. In: CVPR (2023)

    2023

  27. [35]

    In: IEEE/CVF Winter Conf

    Steininger, D., Trondl, A., Croonen, G., Simon, J., Widhalm, V.: The CropAndWeed dataset: A multi-modal learning approach for efficient crop and weed manipulation. In: IEEE/CVF Winter Conf. Applications of Computer Vi- sion (WACV) (2023)

    2023

  28. [36]

    In: IEEE/RSJ Int

    Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS) (2017)

    2017

  29. [37]

    In: CVPR Workshops (2018)

    Tremblay, J., Prakash, A., Acuna, D., Brophy, M., Jampani, V., Anil, C., To, T., Cameracci, E., Boochoon, S., Birchfield, S.: Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In: CVPR Workshops (2018)

    2018

  30. [38]

    ultralytics

    Ultralytics: Ultralytics YOLO11.https : / / docs . ultralytics . com / models / yolo11/(2024)

    2024

  31. [39]

    Vanherle, B., Zoomers, B., Put, J., Van Reeth, F., Michiels, N.: Cut-and- splat: Leveraging Gaussian Splatting for synthetic data generation. In: Int. Conf. Robotics, Computer Vision and Intelligent Systems (ROBOVIS) (2025)

    2025

  32. [40]

    Remote Sensing16(4), 677 (2024)

    Vivoli, E., Bertini, M., Capineri, L.: Deep learning-based real-time detection of surface landmines using optical imaging. Remote Sensing16(4), 677 (2024)

    2024

  33. [41]

    In: CVPR (2023)

    Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR (2023)

    2023

  34. [42]

    sulandterrain field ground terrain, green and dry grass, clover, top-down photograph

    Zanjani, F.G., Abati, D., Wiggers, A., Kalatzis, D., Petersen, J., Cai, H., Habibian, A.: Gaussian Splatting is an effective data generator for 3D object detection. arXiv preprint arXiv:2504.16740 (2025) A Appendix A.1 Configuration Generation.Blender (Cycles) renders at960×96...

    2025 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.