Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Edge-case Synthesis for Fisheye Object Detection: A Data-centric Perspective

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A data-centric pipeline that targets a fisheye detector's measured blind spots with fine-tuned diffusion-generated images, pseudo-labeled by a stronger detector, yields consistent F1 gains without changing the detector architecture.

desk verdict A plausible data-centric pipeline that lacks the control to prove its own central claim, and a final number tuned on test-set pseudo-labels; as a challenge report it's honest, as a research claim it's unsubstantiated. read the letter →

arxiv 2507.16254 v1 pith:ODF5AQAQ submitted 2025-07-22 cs.CV cs.AIcs.LG

Seunghyeon Kim , Kyeongryeol Go This is my paper
classification cs.CVcs.AIcs.LG
keywords fisheyeobjectdetectiondata-centricAIsyntheticdatagenerationtext-to-imagediffusionmodelspseudo-labelingedge-caseanalysisreal-time
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 argues that the fastest route to better fisheye-camera object detection is not a new detector architecture but a data-centric loop: find what the model gets wrong, generate synthetic images that recreate those failure modes, and train on them with pseudo-labels from a stronger detector. Using FishEye8K plus external traffic datasets, the authors analyze the baseline detector's confusions, fine-tune the FLUX.1-dev text-to-image model on fisheye images, craft prompts that push rare classes, small scales, and peripheral distortion, and pseudo-label every synthetic image with Co-DETR. With the lightweight YOLOv11s detector, validation F1 rises from 0.4058 on FishEye8K alone to 0.4907 after all synthetic stages, and a final submission reaches F1 0.5119 on the test split. The authors also report that Co-DETR pseudo-labels outperform the dataset's manual annotations (F1 0.4907 vs 0.4660).

What carries the argument

The load-bearing object is the edge-case synthesis loop. It works in three linked steps: (1) a failure audit that uses confusion matrices, false-positive/false-negative visualizations, and a vision-language model (InternVL3-38B) to tag images with context attributes; (2) targeted generation, where a text-to-image model (FLUX.1-dev), fine-tuned on FishEye8K with a low-rank fine-tuning method (LoKR), is prompted by GPT-4.1-mini to rewrite captions so that rare classes, small scales, and peripheral distortion are overrepresented; and (3) pseudo-labeling, where Co-DETR annotates every synthetic image and unifies labels across real and synthetic training sets. The loop carries the argument because the claimed gains come specifically from aligning synthetic data with the detector's measured blind spots, not from adding arbitrary images.

What would settle it

Train YOLOv11s on FishEye8K plus the same number of synthetic images generated without edge-case targeting, then compare validation F1 against the 0.4907 obtained with the edge-case-targeted sets; if the generic set closes the gap, the improvement is volume, not targeting. A secondary check is to re-evaluate the final model on a human-annotated portion of the test split to see whether thresholds tuned on Co-DETR pseudo-labels reproduce the submitted 0.5119.

Watch

Extended reading notes

Core claim

The central claim is that edge-case synthesis guided by error analysis yields consistent, cumulative performance gains for real-time fisheye object detection without architectural modifications. In the paper's pipeline, a baseline trained only on FishEye8K gives validation F1 0.4058; adding VisDrone and three successive synthetic sets raises it to 0.4218, 0.4682, 0.4839, and 0.4907, and the final model with the validation split included scores F1 0.5119 on the test server. The mechanism is not random data enlargement: each synthetic stage is prompted to re-create specific failure modes (confusing class pairs, small edge-near objects, underrepresented contexts), and the reported class-wise improvements concentrate in previously weak classes such as Pedestrian. The paper also claims pseudo-labels from Co-DETR are not only sufficient but better than the original ground-truth boxes, arguing that label noise and label inconsistency make the stronger detector's unified labeling scheme more effective for training.

Load-bearing premise

The pipeline's success rests on Co-DETR pseudo-labels being accurate enough to train on synthetic images and to stand in for ground truth when picking per-class thresholds on the test set, while the validation gains are attributed to edge-case targeting without a matched-volume control; if the teacher's errors track the student's, or plain data volume drives the gains, the claimed effect is overestimated.

Editorial extensions

If this is right

  • If these results hold, real-time detectors can be improved substantially by data curation alone, without switching to heavier backbones or ensemble inference.
  • Synthetic data generated from a fine-tuned diffusion model can be inserted directly into detection training, provided a reliable teacher detector supplies pseudo-labels.
  • Pseudo-labeling with a strong teacher can beat manual annotations, implying that label consistency across heterogeneous datasets may matter more than label provenance.
  • The pipeline is modular: improving either the text-to-image model or the pseudo-labeler should yield further gains, and no layout-conditioning or annotation filtering is required.

Reading between the lines

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

  • A matched-volume control is missing: the paper does not compare its edge-case-targeted synthetic sets against an equal number of generic synthetic images, so part of the validation gain could be a pure data-volume effect.
  • The final test F1 is obtained after selecting per-class confidence thresholds on Co-DETR pseudo-labels for the test split; if the teacher detector shares the student's systematic errors, the threshold tuning could inflate the reported score relative to a truly blind evaluation.
  • The technique should transfer to other distorted or long-tail perception domains (e.g., wide-angle automotive cameras, low-light surveillance, or rare-class remote sensing), because the mechanism only requires an error audit, a controllable generative model, and a strong pseudo-labeler.
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 presents a data-centric pipeline for fisheye object detection in the AI City Challenge 9 setting. The authors propose to identify model blind spots through error analysis, generate synthetic fisheye-style images with a fine-tuned FLUX.1-dev text-to-image model using prompts designed to target those edge cases, and pseudo-label all real and synthetic data with Co-DETR before training a YOLOv11s detector. Incremental additions of VisDrone and three synthetic versions raise validation F1 from 0.4058 to 0.4907 (Table 1), and the final challenge submission reports a test F1 of 0.5119 after per-class confidence threshold selection. The paper concludes that edge-case synthesis guided by error analysis yields consistent improvements under real-time constraints.

Significance. If the central claim were established, the paper would provide a useful data-centric recipe for improving object detection in a distortion-heavy domain without architectural changes, and it is refreshingly transparent about many implementation details, including prompt construction and the limitation that volume scaling is not characterized. The main contributions are the error-analysis-driven prompt design and the use of fine-tuned diffusion models for fisheye-specific synthesis. However, the reported validation gains are confounded with dataset size, and the headline test F1 is obtained with test-set threshold tuning against pseudo-labels, so the paper does not currently provide a clean measurement of the effect of edge-case targeting. The open acknowledgment of the missing volume-scaling study in the conclusion is commendable but does not substitute for the control.

major comments (3)
  1. [§4.2, Table 1] The central attribution of the gains to edge-case targeting is not supported because every incremental row in Table 1 also increases the number of training images: 5,288 → 13,917 → 22,546 → 39,175 → 49,134 → 51,846. The monotone F1 improvement is exactly what one would expect from additional volume, from Co-DETR pseudo-label distillation, or from the fisheye augmentation applied to VisDrone. To isolate the edge-case synthesis effect, the paper needs a matched-volume control, e.g., training on the same total number of images with synthetic data generated from diversity-oriented prompts only, or with randomly sampled synthetic images that do not target the identified edge cases. Without such a control, the claim that 'edge-case synthesis' drives the improvements is untested.
  2. [§4.2, final submission paragraph] The final test F1 of 0.5119 is not an independent measure of the pipeline. The authors state that class-wise confidence thresholds were selected to maximize F1 on the test split using Co-DETR pseudo-labels as a reference. Because Co-DETR also generates the pseudo-labels used for training all real and synthetic data, the threshold tuning is adapted to the same label source that supervises YOLOv11s. If Co-DETR and YOLOv11s share systematic errors on small or peripheral objects, the reported F1 will overestimate performance relative to a fixed, pre-registered evaluation policy. The paper should either report the server F1 with the default threshold (e.g., 0.5) or determine thresholds on the validation split and report the resulting test F1.
  3. [§4.2, Table 1, 'Final submission' row] The final row of Table 1 shows mAP dropping from 0.4853 to 0.4148 while F1 rises from 0.4907 to 0.5119. This indicates that the F1 gain in the final submission is largely an artifact of per-class threshold selection rather than an improvement in detection quality. Presenting 0.5119 as the outcome of the data-centric pipeline conflates the thresholding effect with the synthesis effect. The paper should report both mAP and F1 under a fixed threshold for the final model, and explicitly separate the contribution of threshold adaptation from the contribution of the training data.
minor comments (5)
  1. [§4.2 / all experiments] All reported F1 and mAP numbers appear to come from a single training run without multiple seeds or error bars; reporting variance over at least three seeds would strengthen the reliability of the incremental comparisons in Table 1.
  2. [§3.3 / supplementary] The text refers to 'details about the prompts' in the supplementary material, but the supplementary provides only a few example prompts and notes that one is 'an example prompt' for Synthetic v3; a complete list of prompt templates and the exact GPT-4.1-mini instruction set would be needed for reproducibility.
  3. [§4.1] The paper estimates an effective inference speed of approximately 16 FPS from the 1280 resolution but never reports an actual measured FPS on the Jetson AGX Orin device, which is central to the challenge's harmonic-mean metric.
  4. [Supplementary, Section 3] There is a typo in the supplementary material: 'lightning' should be 'lighting' in the sentence about background elements.
  5. [§4.3, Table 3a] The augmentation ablations in Table 3a are performed on the 640-resolution FishEye8K-only setting, while the fisheye augmentation in Table 3b is evaluated with all datasets; the text should clarify that these two ablations are not directly comparable.

Circularity Check

2 steps flagged · score 6.0 of 10

Test-set F1 is tuned to Co-DETR pseudo-labels and the target attribution is entangled with data volume, but the core pipeline is otherwise independently evaluated.

  1. fitted input called prediction [Section 4.2, paragraph on final submission]
    "For final submission, we applied class-wise optimal confidence thresholds to filter predictions, aiming to maximize the F1 score on the test set. These thresholds were determined based on Co-DETR pseudo labels for the test split, which served as a reference to approximate the true distribution in the absence of ground-truth annotation."

    The reported test-set result (F1 0.5119) is not the output of a fixed inference policy. The per-class thresholds are fitted to maximize F1 against Co-DETR pseudo-labels on the test images themselves, so the score reflects adaptation to the pseudo-labeler's outputs rather than an independent measurement of detector quality on the test split. If Co-DETR and YOLOv11s share systematic errors (e.g., on small peripheral objects), threshold fitting on pseudo-labels partly optimizes the metric against the labeler's own mistakes, making this evaluation step circular with respect to the pseudo-labeler used throughout training.

  2. fitted input called prediction [Section 4.2, Table 1 caption and preceding paragraph; Conclusion Section 5]
    "Incremental addition of the training data led to consistent performance improvements, as evidenced by Tab. 1. ... These findings underscore the efficacy of targeted data synthesis in addressing rare or challenging categories."

    The central attribution claim ('edge-case synthesis yields the gains') is statistically forced by construction of the experiment: every '+Synthetic' row adds tens of thousands of images (5,288 to 49,134 to 51,846 total), so the monotone F1 increases co-vary with volume, and also with Fisheye-8K fine-tuning of FLUX and Co-DETR pseudo-label distillation. No matched-volume control isolates the edge-case prompts of v2/v3 from plain volume expansion. The paper itself concedes this gap in the conclusion ('It is also important to characterize how performance scales with the volume of generated data'). The attribution of Table 1 to edge-case targeting is therefore a fitted-input-as-prediction pattern: the input being varied includes a confound that predicts the observed trend by itself.

full rationale

Most of the paper's pipeline is not circular: the validation-split results in Table 1 use held-out ground-truth annotations (the paper explicitly states ground truth is used only for evaluation there), the generative-model fine-tuning is a standard external procedure, and the data-augmentation ablations are internally controlled. However, two load-bearing evaluation steps reduce to their own inputs. First, the final submitted test F1 (0.5119) is obtained after fitting per-class confidence thresholds to maximize F1 against Co-DETR pseudo-labels on the test split itself; because Co-DETR also generates all training labels, the reported test number is at least partly a function of the pseudo-labeler's predictions rather than an independent measure. Second, the core claim that targeted edge-case synthesis drives the gains is confounded by volume: each incremental row in Table 1 adds large numbers of images, so the monotone improvement is exactly what volume or pseudo-label distillation would predict, and no ablation holds image count fixed while varying edge-case prompts. The paper's own conclusion admits the missing volume-scaling control. Together these make the central reported result partially circular with respect to the pseudo-labeler and the un-controlled volume confound, though substantial non-circular components (validation ground-truth evaluation, qualitative comparisons, augmentation ablations) remain; score 6 per the rubric.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central claim depends on the accuracy of pseudo-labels and on several hand-chosen quantities: synthetic data volumes, the fisheye augmentation distribution, and the test-set confidence thresholds used for the final submission. The thresholds are the most serious because they directly influence the reported F1.

free parameters (3)
  • class-wise confidence thresholds on test split = not reported
    Selected to maximize F1 against Co-DETR pseudo-labels on the test set; used in the final submission F1 of 0.5119.
  • synthetic data volumes per stage = v1=8,629; v2=16,629; v3=9,959 images
    Hand-chosen amounts of synthetic data; the incremental gains in Table 1 are not controlled for volume.
  • fisheye augmentation distortion coefficients and focal length ranges = not specified
    Randomly sampled in a 'plausible range' for VisDrone augmentation; parameterization not disclosed.
assumptions (4)
  • domain assumption Co-DETR pseudo-labels on synthetic and real images are accurate enough to serve as supervision and to approximate test ground truth for threshold tuning.
    Invoked in §3.3 and §4.2; no verification against human labels on synthetic images.
  • domain assumption Fine-tuning FLUX.1-dev on FishEye8K with VLM captions produces images sharing the relevant fisheye distortion and scene statistics.
    Invoked in §3.3 and Figure 3; only qualitatively demonstrated.
  • domain assumption Validation error analysis identifies failure modes that generalize to the test distribution.
    Invoked in §3.2; the test set is never used during analysis, so transfer is assumed.
  • domain assumption Per-class thresholds tuned on test pseudo-labels do not constitute training or biased evaluation.
    The paper asserts this in §4.2, but tuning on the test split is a form of adaptation and can inflate F1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Edge-case Synthesis for Fisheye Object Detection: A Data-centric Perspective." pith.science (2026). https://pith.science/paper/ODF5AQAQ

@misc{pith2026250716254,
  author       = {Pith},
  title        = {Pith review of: Edge-case Synthesis for Fisheye Object Detection: A Data-centric Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODF5AQAQ}},
  note         = {Machine review of arXiv:2507.16254}
}
read the original abstract

Fisheye cameras introduce significant distortion and pose unique challenges to object detection models trained on conventional datasets. In this work, we propose a data-centric pipeline that systematically improves detection performance by focusing on the key question of identifying the blind spots of the model. Through detailed error analysis, we identify critical edge-cases such as confusing class pairs, peripheral distortions, and underrepresented contexts. Then we directly address them through edge-case synthesis. We fine-tuned an image generative model and guided it with carefully crafted prompts to produce images that replicate real-world failure modes. These synthetic images are pseudo-labeled using a high-quality detector and integrated into training. Our approach results in consistent performance gains, highlighting how deeply understanding data and selectively fixing its weaknesses can be impactful in specialized domains like fisheye object detection.

Figures

Figures reproduced from arXiv: 2507.16254 by the authors.

Figure 1
Figure 1. A simple diagram of our proposed pipeline. In addition to the dataset provided by the challenge, we incorporate external sources [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. (a): An image generated with a pre-trained generative [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Radar plot illustrating class-wise AP across different [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison of false positives and false negatives as more synthetic data is added. False positives are marked in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 1
Figure 1. Figure 1: Qualitative comparison of synthetic images generated by the pre-trained diffusion model (top) and the fine-tuned model (bot [PITH_FULL_IMAGE:figures/full_fig_p012_1.png]
Figure 2
Figure 2. Figure 2: Comparison of model predictions on validation images using synthetic data v1 (left), v2 (center), and v3 (right). False positives [PITH_FULL_IMAGE:figures/full_fig_p013_2.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Physics-aware Masked Diffusion-based Flood Simulation for Urban Fisheye Disaster Detection

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A diffusion-based inpainting system that synthesizes water-level-controlled flood scenes from single fisheye images for disaster detection training.

  2. Towards Continual Expansion of Data Coverage: Automatic Text-guided Edge-case Synthesis

    cs.CV 2025-09 unverdicted novelty 6.0 of 10

    Automated LLM-based prompt engineering for text-to-image edge-case synthesis improves object detection robustness on the FishEye8K benchmark over naive augmentation and manual prompts.

Reference graph

Works this paper leans on

55 extracted references · 34 canonical work pages · cited by 2 Pith papers

  1. [1]

    Slicing aided hyper inference and fine-tuning for small object detection

    Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel. Slicing aided hyper inference and fine-tuning for small object detection. In 2022 IEEE international con- ference on image processing (ICIP) , pages 966–970. IEEE,

    work page 2022

  2. [2]

    Synthetic data from diffusion models improves imagenet classification

    Shekoofeh Azizi, Simon Kornblith, Chitwan Saharia, Mo- hammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023. 3

    arXiv 2023

  3. [3]

    Rain removal in traffic surveillance: Does it matter? IEEE Transactions on Intelligent Transportation Systems, 20(8):2802–2819, 2018

    Chris H Bahnsen and Thomas B Moeslund. Rain removal in traffic surveillance: Does it matter? IEEE Transactions on Intelligent Transportation Systems, 20(8):2802–2819, 2018. 3

    work page 2018

  4. [4]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 2

    2020

  5. [5]

    Geodiffusion: Text- prompted geometric control for object detection data gen- eration

    Kai Chen, Enze Xie, Zhe Chen, Yibo Wang, Lanqing Hong, Zhenguo Li, and Dit-Yan Yeung. Geodiffusion: Text- prompted geometric control for object detection data gen- eration. arXiv preprint arXiv:2306.04607, 2023. 3

    arXiv 2023

  6. [6]

    Region-aware text-to-image generation via hard binding and soft refinement

    Zhennan Chen, Yajie Li, Haofan Wang, Zhibo Chen, Zhengkai Jiang, Jun Li, Qian Wang, Jian Yang, and Ying Tai. Region-aware text-to-image generation via hard binding and soft refinement. arXiv preprint arXiv:2411.06558, 2024. 3

    arXiv 2024

  7. [7]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 3

    2009

  8. [8]

    The unmanned aerial vehicle benchmark: Object detection and tracking

    Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The unmanned aerial vehicle benchmark: Object detection and tracking. In Proceedings of the European con- ference on computer vision (ECCV) , pages 370–386, 2018. 3

    work page 2018

Show all 55 references
  1. [9]

    Visdrone-det2019: The vision meets drone ob- ject detection in image challenge results

    Dawei Du, Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Lin, Qinghua Hu, Tao Peng, Jiayu Zheng, Xinyao Wang, Yue Zhang, et al. Visdrone-det2019: The vision meets drone ob- ject detection in image challenge results. In Proceedings of the IEEE/CVF international conference on compu...

    2019

  2. [10]

    Robust data augmen- tation and ensemble method for object detection in fisheye camera images

    Viet Hung Duong, Duc Quyen Nguyen, Thien Van Luong, Huan Vu, and Tien Cuong Nguyen. Robust data augmen- tation and ensemble method for object detection in fisheye camera images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7017– ...

    2024

  3. [11]

    Sim- ple copy-paste is a strong data augmentation method for in- stance segmentation

    Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung- Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Sim- ple copy-paste is a strong data augmentation method for in- stance segmentation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recogni...

    2021

  4. [12]

    Fast r-cnn

    Ross Girshick. Fast r-cnn. In Proceedings of the IEEE inter- national conference on computer vision , pages 1440–1448,

  5. [13]

    Fisheye8k: A benchmark and dataset for fisheye camera object detection

    Munkhjargal Gochoo, Munkh-Erdene Otgonbold, Erkhem- bayar Ganbold, Jun-Wei Hsieh, Ming-Ching Chang, Ping- 7 Yang Chen, Byambaa Dorj, Hamad Al Jassmi, Ganzorig Bat- nasan, Fady Alnajjar, et al. Fisheye8k: A benchmark and dataset for fisheye camera object detection. In Proceedin...

    2023

  6. [14]

    Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022

    Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 3

    2022 arXiv

  7. [15]

    Ultralytics yolo11, 2024

    Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 5

    2024

  8. [16]

    Ultralytics YOLO, 2023

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. Ultralytics YOLO, 2023. 2, 4

    2023

  9. [17]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 3, 4, 5

    2024

  10. [18]

    Gligen: Open-set grounded text-to-image generation

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jian- wei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511–22521, 2023. 3

    2023

  11. [19]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceeding...

    2014

  12. [20]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision , pages 38–55. Springe...

    2024

  13. [21]

    Rt-detrv2: Improved base- line with bag-of-freebies for real-time detection transformer

    Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu. Rt-detrv2: Improved base- line with bag-of-freebies for real-time detection transformer. arXiv preprint arXiv:2407.17140, 2024. 2

    2024 arXiv

  14. [22]

    Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation

    Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic data generation for pixel-level semantic segmentation. Advances in Neural Information Processing Systems, 36:76872–76892, 2023. 3

    2023

  15. [23]

    Gpt-4.1, 2024

    OpenAI. Gpt-4.1, 2024. Accessed: 2025-07-05. 4

    2024

  16. [24]

    D-fine: redefine regression task in de- trs as fine-grained distribution refinement

    Yansong Peng, Hebei Li, Peixi Wu, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. D-fine: redefine regression task in de- trs as fine-grained distribution refinement. arXiv preprint arXiv:2410.13842, 2024. 2

    2024 arXiv

  17. [25]

    Improving object detection to fisheye cameras with open-vocabulary pseudo- label approach

    Long Hoang Pham, Quoc Pham-Nam Ho, Duong Nguyen- Ngoc Tran, Tai Huu-Phuong Tran, Huy-Hung Nguyen, Duong Khac Vu, Chi Dai Tran, Ngoc Doan-Minh Huynh, Hyung-Min Jeon, Hyung-Joon Jeon, et al. Improving object detection to fisheye cameras with open-vocabulary pseudo- label approac...

    2024

  18. [26]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4

    2024 arXiv

  19. [27]

    You only look once: Unified, real-time object de- tection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object de- tection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016. 2

    2016

  20. [28]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 3

    2022

  21. [29]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  22. [30]

    Better aggregation in test-time augmentation

    Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1214–1223, 2021. 3

    2021

  23. [31]

    Road object detection robust to distorted objects at the edge regions of images

    Wooksu Shin, Donghyuk Choi, Hancheol Park, and Jeongho Kim. Road object detection robust to distorted objects at the edge regions of images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7245–7251, 2024. 2

    2024

  24. [32]

    Weighted boxes fusion: Ensembling boxes from different ob- ject detection models

    Roman Solovyev, Weimin Wang, and Tatiana Gabruseva. Weighted boxes fusion: Ensembling boxes from different ob- ject detection models. Image and Vision Computing , 107: 104117, 2021. 2

    2021

  25. [33]

    Aerogen: Enhancing remote sensing object detection with diffusion-driven data generation

    Datao Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, Dongsheng Jiang, Yin Li, and Deyu Meng. Aerogen: Enhancing remote sensing object detection with diffusion-driven data generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages...

    2025

  26. [34]

    You only learn one rep- resentation: Unified network for multiple tasks

    CY Wang, IH Yeh, and HYM Liao. You only learn one rep- resentation: Unified network for multiple tasks. arxiv 2021. arXiv preprint arXiv:2105.04206, 2021. 2, 3

    2021 arXiv

  27. [35]

    Yolov9: Learning what you want to learn using pro- grammable gradient information

    Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using pro- grammable gradient information. In European conference on computer vision, pages 1–21. Springer, 2024. 2

    2024

  28. [36]

    Exploiting diffusion prior for real-world image super-resolution

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929–5949, 2024. 2

    2024

  29. [37]

    The 8th ai city challenge

    Shuo Wang, David C Anastasiu, Zheng Tang, Ming- Ching Chang, Yue Yao, Liang Zheng, Mohammed Shaiqur Rahman, Meenakshi S Arya, Anuj Sharma, Pranamesh Chakraborty, et al. The 8th ai city challenge. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2024

  30. [38]

    Internimage: Exploring large-scale vi- sion foundation models with deformable convolutions

    Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vi- sion foundation models with deformable convolutions. In Proceedings of the IEEE/CVF conference on computer vi- si...

    2023

  31. [39]

    Instancediffusion: Instance- level control for image generation

    Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Ro- hit Girdhar, and Ishan Misra. Instancediffusion: Instance- level control for image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6232–6242, 2024. 3

    2024

  32. [40]

    Ua-detrac: A new benchmark and protocol for multi-object detection and tracking

    Longyin Wen, Dawei Du, Zhaowei Cai, Zhen Lei, Ming- Ching Chang, Honggang Qi, Jongwoo Lim, Ming-Hsuan Yang, and Siwei Lyu. Ua-detrac: A new benchmark and protocol for multi-object detection and tracking. Computer Vision and Image Understanding, 193:102907, 2020. 3

    2020

  33. [41]

    Datasetdm: Synthesizing data with perception annota- tions using diffusion models

    Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annota- tions using diffusion models. Advances in Neural Informa- tion Processing Systems, 36:54683–54695, 2023. 3

    2023

  34. [42]

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 1206–1217,

  35. [43]

    A comprehensive overview of fish-eye camera distortion correction methods

    Jian Xu, De-Wei Han, Kang Li, Jun-Jie Li, and Zhao-Yuan Ma. A comprehensive overview of fish-eye camera distortion correction methods. arXiv preprint arXiv:2401.00442, 2023. 1

    2023 arXiv

  36. [44]

    Reco: Region-controlled text-to-image genera- tion

    Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image genera- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 14246–14255,

  37. [45]

    Navigating text- to-image customization: From lycoris fine-tuning to model evaluation

    Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text- to-image customization: From lycoris fine-tuning to model evaluation. In The Twelfth International Conference on Learning Representations, 2023. 5

    2023

  38. [46]

    Detrs beat yolos on real-time object detection

    Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16965–16974, 2024. 2

    2024

  39. [47]

    Migc: Multi-instance generation controller for text-to-image synthesis

    Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Migc: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6818–6828,

  40. [48]

    Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image models

    Dewei Zhou, Mingwei Li, Zongxin Yang, and Yi Yang. Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image models. arXiv preprint arXiv:2503.12885, 2025. 3

    2025 arXiv

  41. [49]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shen- glong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. 4, 5

    2025 arXiv

  42. [50]

    Unpaired image-to-image translation using cycle- consistent adversarial networks

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle- consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision , pages 2223– 2232, 2017. 3

    2017

  43. [51]

    Detrs with col- laborative hybrid assignments training

    Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with col- laborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748–6758, 2023. 2, 4 9 Supplementary Material for Edge-case Synthesis for Fisheye Object Detection...

    2023

  44. [52]

    We specifically guided the model to ensure that the base captions included the key objects targeted in the fisheye dataset

    Prompt engineering for base captions In this section, we provide the prompt used with the InternVL3-38B model to obtain the base captions. We specifically guided the model to ensure that the base captions included the key objects targeted in the fisheye dataset. The captions o...

  45. [53]

    These prompts were given to GPT-4.1-mini to rewrite base captions that are originally extracted from real data so that they better describe the types of images we aimed to generate

    Prompt engineering for rephrased captions In this section, we provide the detailed prompts used in our pipeline. These prompts were given to GPT-4.1-mini to rewrite base captions that are originally extracted from real data so that they better describe the types of images we a...

  46. [54]

    Figure 1 presents five representative examples, where each column compares images generated from the base pre-trained model (top) and our fine-tuned model (bottom)

    Fine-tuned synthetic image quality To better illustrate the effectiveness of our data generation pipeline, we provide qualitative examples of synthetic images generated before and after fine-tuning the diffusion model. Figure 1 presents five representative examples, where each...

  47. [55]

    Figure 2 shows that as the synthetic data improves, both false positives and false negatives are gradually reduced, indicating more accurate and confident detection

    Effect of synthetic data on prediction quality To further demonstrate the effectiveness of our synthetic data in improving detection performance, we visualize model pre- dictions on three different validation images under incremental incorporation of synthetic datasets (v1, v2...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.