Resmi CarDD test seti: makaleler ve bizOfficial CarDD test split: papers vs. us
Aynı 374 fotoğraf, aynı COCO ölçütü. Mavi = telefonda çalışan modelimiz; sarı = Hasar AI 2.1 dikkatli mod (aynalı 2. geçiş + Weighted Boxes Fusion); gri = makalelerde yayımlanan sonuçlar.Same 374 photos, same COCO metric. Blue = our on-phone model; amber = Hasar AI 2.1 careful mode (mirrored 2nd pass + Weighted Boxes Fusion); grey = published results.
Kutu AP (IoU 0,50–0,95)Box AP (IoU 0.50–0.95)
Kutu AP50Box AP50
| Hasar türüDamage type | AP · hızlıfast | AP50 · hızlıfast | AP · 2.1 dikkatlicareful | AP50 · 2.1 | DCN+ R101 (maske AP, makale)(mask AP, paper) |
|---|---|---|---|---|---|
| GöçükDent | 33.0 | 58.4 | 37.8 | 64.3 | 40.5 |
| ÇizikScratch | 38.1 | 66.5 | 42.1 | 69.9 | 34.3 |
| ÇatlakCrack | 31.7 | 52.1 | 32.3 | 53.8 | 16.6 |
| Cam kırığıGlass shatter | 93.2 | 97.2 | 93.4 | 97.3 | 89.6 |
| Kırık lambaBroken lamp | 76.9 | 88.3 | 80.9 | 91.1 | 70.8 |
| Patlak lastikFlat tyre | 86.8 | 91.4 | 87.0 | 91.9 | 90.0 |
Küçük nesne AP: 24.9 → 31.5 (WBF ile) · makalenin DCN+ R101 kutu APs değeri 37,1. Ortalama süre 496 ms/fotoğraf (CPU, int8).Small-object AP: 24.9 → 31.5 (with WBF) · the paper's DCN+ R101 box APs is 37.1. Mean 496 ms/photo (CPU, int8).
Dürüst okuma:Honest reading: (1) Temel ağırlıklarımız (shanexf YOLO11m) CarDD eğitim setiyle eğitildi; test setinin eğitimde kullanılmadığını varsayıyoruz ama doğrulayamıyoruz. (2) Model kartındaki “AP50 0,899” iddiasını yeniden üretemedik: int8 telefon modelinde 75,7 ölçtük. (3) ¹ YOLOv11 ve YOLOv9 makaleleri farklı bir bölünme kullanmış olabilir. (4) Makale sınıf değerleri maske AP'sidir, bizimkiler kutu AP'si; doğrudan karşılaştırılamaz. (5) CarDD yalnızca araştırma lisanslıdır; ticari ürün için lisans-temiz yeniden eğitim (v3) gerekir.(1) Our base weights (shanexf YOLO11m) were trained on CarDD train; we assume the test split was not used in training but cannot verify it. (2) We could not reproduce the model card's “AP50 0.899”: we measure 75.7 for the int8 phone model. (3) ¹ The YOLOv11 and YOLOv9 papers may use a different split. (4) The paper's per-class values are mask AP, ours are box AP; not directly comparable. (5) CarDD is research-only; a commercial product needs a licence-clean retrain (v3).
Diğer makalelerden çıkanlarWhat other papers tell us
| MakalePaper | BildirilenReported | Çekicin için anlamıWhat it means for Çekicin |
|---|---|---|
| Grounding Agentic VLMs with Dedicated Segmentation (arXiv 2608.02470, 2026) | Qwen-VL damage-type accuracy 87.3%; hallucination: text-only 92% → image 78% → segmentation-grounded 31% | our cloud 2nd opinion is detector-grounded (same idea); not benchmarked yet |
| INS-MMBench (ICCV 2025) | GPT-4o auto-insurance score 86.0; damage-severity ≈30% | VLMs are weak at severity → keep severity rule-based + operator |
| WildRoadBench (2026) | best VLM 42.1% AP50 grounding | VLMs cannot replace a detector for localisation |
| Small Dents, Big Impact (arXiv 2508.15431) | dent detector 432 TP / 424 FP | dent false alarms are a known problem; our decision layer cut dent FA |
| CarDD paper pre-filter | VGG16 damaged/undamaged precision 91.6%, recall 97.6% | we use a calibrated gate (AUROC .874) instead; CLIP gate = next experiment |
| MVA-CDD multi-view (TITS 2025, CDDM 4414 pairs) | close-up + distant fusion beats single view | multi-photo aggregation added to the roadmap / API |
| RF-DETR (ICLR 2026) | RF100-VL AP50: M 87.4 vs D-FINE-M 85.5; Apache-2.0 | licence-clean retraining route for v3 (needs GPU) |
v2.2: CarDD verisiyle eğitim denemesi (CPU)v2.2: training on CarDD data (CPU)
Ağın kendisini GPU olmadan yeniden eğitemeyiz; bu yüzden donmuş dedektörün üstüne CarDD ile eğitilen iki aşama denedik: öğrenilmiş yeniden puanlayıcı (gradient boosting, 44 özellik) ve 2×2 döşemeli çıkarım (SAHI). Kural: val ile eğit/seç, test’e bir kez bak. Tüm koşular Weights & Biases’ta.We cannot retrain the network itself without a GPU, so we trained two stages with CarDD on top of the frozen detector: a learned rescorer (gradient boosting, 44 features) and 2×2 tiled inference (SAHI). Rule: fit/select on val, look at test once. All runs are on Weights & Biases.
Ezber kontrolü (aynı model, aynı ölçüt)Memorisation check (same model, same metric)
| CarDD split | AP | AP50 |
|---|---|---|
| train (250 img) | 79.4 | 92.9 |
| val (810) | 64.8 | 79.2 |
| test (374) | 59.9 | 75.6 |
Model eğitim görüntülerini ezberlemiş: train tahminleriyle ikinci aşama eğitmek yanıltıcı olur. Model kartındaki “AP50 0,899” de yazarın kendi CarDD bölünmesinde ölçülmüş; resmi test ile karşılaştırılamaz.The model has memorised the training images, so training a second stage on train predictions would mislead. The model card’s “AP50 0.899” was measured on the author’s own CarDD re-split and is not comparable with the official test.
Test sonucu (resmi 374 fotoğraf)Test result (official 374 photos)
| AyarSetting | AP | AP50 | APs | ΔAP [95% CI] |
|---|---|---|---|---|
| hızlıfast | 59.93 | 75.62 | 25.15 | |
| + yeniden puanlayıcırescorer | 60.08 | 75.90 | 25.24 | +0.16 [+0.00, +0.33] |
| dikkatli (2.1)careful (2.1) | 62.23 | 77.96 | 31.61 | |
| + yeniden puanlayıcırescorer | 62.47 | 78.09 | 31.72 | +0.23 [+0.03, +0.42] |
| + döşemelertiles (SAHI) | 62.17 | 77.83 | 31.27 | -0.07 [-0.14, +0.05] |
Sonuç: CPU ile eğitilen ek aşamalar en fazla birkaç onda bir puan kazandırıyor. Gerçek kazanç için ağın GPU’da yeniden eğitilmesi gerekiyor (aşağıdaki not defteri, W&B’ye otomatik kayıt).Conclusion: CPU-trained extra stages gain at most a few tenths of a point. Real gains need GPU retraining of the network (notebook below, logs to W&B automatically).
Makalelerden 100 iyileştirme becerisi100 improvement skills from the papers
Her beceri, kaynak makale numaralarıyla. Filtreleyin:Each skill with its source paper numbers. Filter:
- 2.1’de yapıldıdone in 2.11. Evaluate on the official CarDD test split (374 imgs, COCO box AP) so numbers are comparable with papers #9 #1 #3
- 2.1’de yapıldıdone in 2.12. Report AP50 AND AP50:95 AND per-class AP; never a single accuracy number #9 #1
- varhave3. Report image-level recall + false-alarm rate on intact cars next to box AP (papers rarely do) #9 #11
- varhave4. Keep an out-of-domain test (VehiDE/AutoDamageIQ) besides in-domain CarDD #76 #8
- varhave5. Include intact close-ups as negatives (dealer photos are too easy) #11 #9
- varhave6. Use a damaged-vs-intact filter when collecting data (CarDD VGG16 filter: P 91.6 / R 97.6) #9
- sırada (CPU)next (CPU)7. Remove near-duplicates across splits (Duplicate Cleaner) #9
- varhave8. Bootstrap 95% CIs for every headline metric #8
- 2.1’de yapıldıdone in 2.19. Small/medium/large AP: small objects are where car-damage models fail (CarDD APs 37 vs APl 66) #9
- sırada (CPU)next (CPU)10. Benchmark VLMs with structured (type, part, severity) tuples, not free text #5 #10
- sırada (CPU)next (CPU)11. Separate hallucination rate for VLM answers (text-only 92% → grounded 31%) #2
- varhave12. Track every version on the same frozen sets (trend page) #8
- varhave13. Hard set with lamp/tyre and night/blur photos #3 #11
- sırada (CPU)next (CPU)14. Collect consented Çekicin photos (KVKK) into a private dataset for retraining #28 #33
- sırada (CPU)next (CPU)15. Label severity + part, not only type (insurer-style data) #25 #39 #73
- varhave16. YOLO11m is the best speed/accuracy trade-off on CarDD (63.3 AP50:95 vs RT-DETR-L) #1 #4
- v3 · GPUv3 · GPU17. Licence-clean retrain on RF-DETR-M / D-FINE (Apache-2.0) #12
- v3 · GPUv3 · GPU18. Multi-scale training (CarDD: +2.0 AP) #9 #3
- v3 · GPUv3 · GPU19. Focal loss for hard classes crack/scratch (CarDD: +4.1 AP) #9
- v3 · GPUv3 · GPU20. Deformable convolutions for irregular damage shapes (DCN+: best in CarDD paper) #9 #35
- v3 · GPUv3 · GPU21. Attention modules (CBAM/SE/coordinate) in the YOLO neck #4 #49 #23
- v3 · GPUv3 · GPU22. Heterogeneous / AKConv kernels for thin scratches #49 #23
- v3 · GPUv3 · GPU23. NMS-free heads (YOLOv10) for lower latency #23
- v3 · GPUv3 · GPU24. Dilated-conv FPN for multi-scale damage (DamageNet) #35
- v3 · GPUv3 · GPU25. Diffusion-based detectors (DiffusionDet, C-DiffDet+) with scene context #32 #90
- v3 · GPUv3 · GPU26. Region-aware mixture-of-experts for fine-grained classes #16
- v3 · GPUv3 · GPU27. Instance segmentation (mask) for area-based severity #18 #35 #68
- v3 · GPUv3 · GPU28. IoU-loss variants (CIoU/WIoU) for small boxes #4 #49
- v3 · GPUv3 · GPU29. Synthetic / GAN-augmented damage images for rare classes #47 #52
- v3 · GPUv3 · GPU30. Hyper-parameter search (e.g. Harris-Hawks-tuned YOLOv8-seg) #53
- v3 · GPUv3 · GPU31. Knowledge distillation m → n for phones #23 #49
- varhave32. Train with intact close-ups as background images (hard negatives) #11
- v3 · GPUv3 · GPU33. Copy-paste augmentation of damage onto intact panels #47
- v3 · GPUv3 · GPU34. Self-/semi-supervised pre-training on unlabeled claim photos #67
- v3 · GPUv3 · GPU35. Quantisation-aware training to keep int8 accuracy #23
- varhave36. Low detector threshold (0.05) + learned decision layer instead of a fixed 0.25 #11
- varhave37. Horizontal-flip test-time augmentation #3
- 2.1’de yapıldıdone in 2.138. Weighted Boxes Fusion of TTA views instead of NMS (better boxes) #1 #62
- 2.1’de yapıldıdone in 2.139. Evaluate at the COCO protocol (conf 0.001, top 300) AND at the shipped threshold #9
- varhave40. Class-aware NMS 0.45 (browser = Python parity) #1
- varhave41. Letterbox exactly like training (resize filter matters: measured) #8
- sırada (CPU)next (CPU)42. Tiled / zoomed inference for small damage in wide shots (SAHI-style) #9 #49
- varhave43. Ask for a close-up when the car is small in the frame (wide crash scenes fail) #7 #9
- varhave44. Photo-quality gate: blur (Laplacian) + darkness → retake #3 #59
- sırada (CPU)next (CPU)45. Glare/reflection check before accepting scratch/dent #11 #2
- varhave46. Ensemble the damage model with the parts model (features) #62 #21
- sırada (CPU)next (CPU)47. Visual-embedding (CLIP) score as an extra decision feature #62 #31
- varhave48. Damaged-vs-intact global classifier as a pre-filter feature #9 #21
- varhave49. int8 quantised ONNX + WASM for on-device privacy #23 #64
- varhave50. Cache models in the browser; never upload the photo #64
- sırada (CPU)next (CPU)51. Multi-scale TTA (640 + 800 px) for small damage #9 #49
- sırada (CPU)next (CPU)52. Box-level ensemble of two damage detectors fused with WBF #1 #62
- varhave53. Calibrated logistic stacker on 34 features (portable to JS/JSON) #8
- varhave54. Per-type heads with thresholds chosen inside CV folds only #8
- varhave55. False-alarm budget as an explicit constraint #11
- varhave56. Safety floor: lamp/tyre ≥ 0.60 always shown (tow-relevant) #3
- varhave57. Grouped CV so crops of one car never leak across folds #8
- varhave58. Confidence levels (high/medium/low) instead of raw scores in the UI #5
- varhave59. Abstain ("retake") instead of guessing on bad photos #2
- sırada (CPU)next (CPU)60. Temperature/Platt check of the reported confidences (reliability diagram) #8
- sırada (CPU)next (CPU)61. Per-class operating points tuned for business cost (missed tow ≫ false alarm) #25
- sırada (CPU)next (CPU)62. Monitor drift: weekly score histograms on real traffic #28
- varhave63. Report AUROC damaged-vs-intact for the gate #9 #21
- varhave64. Damage + part localisation together (which part is damaged) #17 #43 #98
- 2.1’de yapıldıdone in 2.165. Severity index from damage area / part area (DSI) #3 #25
- varhave66. Severity → action mapping: drivable / workshop / tow #25 #39
- sırada (CPU)next (CPU)67. Three-quarter-view standard photos for consistent assessment #43
- sırada (CPU)next (CPU)68. Repair-cost estimation from (type, part, severity) #13 #15 #26
- sırada (CPU)next (CPU)69. Ontology / rule reasoning for cost (parts catalogue) #82
- varhave70. Glass damage pipeline separate from body damage #59
- sırada (CPU)next (CPU)71. Tyre: flat vs. surface damage distinction #84
- varhave72. Count instances per type (2 areas of crack) in the result #17
- varhave73. Show boxes on the photo for transparency (explainability) #54
- sırada (CPU)next (CPU)74. Damage-area / part-area ratio for repair-vs-replace hints #25 #3
- varhave75. Ground the VLM on detector output (hypotheses to verify) → fewer hallucinations #2 #50
- varhave76. VLM only as a second opinion, never as the detector (weak localisation) #5 #10
- sırada (CPU)next (CPU)77. Structured JSON answers from the VLM (type, part, severity, confidence) #5
- sırada (CPU)next (CPU)78. Text-guided phrase grounding for user questions ("is the bumper cracked?") #31
- sırada (CPU)next (CPU)79. Spatially-aware VLM prompting (regions, coordinates) #50
- sırada (CPU)next (CPU)80. LLM-generated claim report from detections #93 #98
- sırada (CPU)next (CPU)81. Fraud signals (re-used / edited photos) before claims #46 #93
- varhave82. Language-specific prompts (TR/EN) and answer checks #5
- varhave83. Voice questions answered by the assistant (Telegram) #98
- varhave84. Keep the human operator in the loop for the final decision #10 #5
- 2.1’de yapıldıdone in 2.185. Multi-photo aggregation per incident (close-up + distant) #7
- sırada (CPU)next (CPU)86. Close-up/distant pair alignment (patch-to-patch) #67 #7
- sırada (CPU)next (CPU)87. Track damage across video frames (MOT) for walk-around videos #87
- v3 · GPUv3 · GPU88. 3D lifting of damage (Gaussian splatting) for big repairs #7
- sırada (CPU)next (CPU)89. Guided capture: ask for missing views (front/side/rear) #43 #7
- sırada (CPU)next (CPU)90. De-duplicate the same damage seen in two photos of one incident #7 #67
- varhave91. Real-time on phone (≤ 2–3 s) as in lightweight YOLO papers #23 #49 #64
- varhave92. Same decision layer on device, API, Telegram, WhatsApp (one JSON) #28
- varhave93. Public wiki + changelog for every version #8
- varhave94. Model cards with licence and training data (non-commercial CarDD risk) #9 #12
- varhave95. Clear disclaimer: pre-assessment, not an expert report #10
- varhave96. Explain WHY (advice text per type) not only WHAT #54
- varhave97. One-tap next action (call tow / workshop) #28 #40
- varhave98. A/B test new versions on the frozen benchmarks before shipping #8
- varhave99. Privacy: KVKK consent + no upload by default #28
- 2.1’de yapıldıdone in 2.1100. Publish a paper-style comparison page so partners can check claims #8 #54
100 makalelik veritabanıThe 100-paper database
OpenAlex + arXiv'den 2.400+ adaydan seçildi (konu filtresi + atıf + güncellik). Model, veri seti ve özetteki sayılar otomatik çıkarıldı; ilk 12 makale elle kontrol edildi. En sık modeller: Transfer learning (25), YOLOv8 (15), Mask R-CNN (15), ResNet (11), VGG (8), YOLO (7).Selected from 2,400+ candidates from OpenAlex + arXiv (topic filter + citations + recency). Models, datasets and numbers in the abstract are extracted automatically; the first 12 papers were checked by hand. Most frequent models: Transfer learning (25), YOLOv8 (15), Mask R-CNN (15), ResNet (11), VGG (8), YOLO (7).
| # | YılYear | MakalePaper | ModellerModels | VeriData | Özetteki sonuçResult in abstract |
|---|---|---|---|---|---|
| 1 | 2025 | Car Damage Detection using YOLOv11 and RT-DETRAzmi Rizqullah Rabbani, Gamma Kosala · International Seminar on Research of Information Technology | YOLOv11, YOLOv10, RT-DETR, DETR | CarDD | accuracy 63.34% |
| 2 | 2026 | Grounding Agentic VLMs with Dedicated Segmentation for Fine-Grained Vehicle Damage AssessmentVishwajeet Shivaji Hogale, Anjali Pai, Nitya Ravi · arXiv | Qwen-VL, VLM/MLLM | accuracy 87.3% | |
| 3 | 2025 | Smart Car Damage Assessment Using Enhanced YOLO Algorithm and Image Processing TechniquesMuhammad Remzy Syah Ramazhan, Alhadi Bustamam, Rinaldi Anwar Buyung · Information | YOLOv9 | CarDD | |
| 4 | 2025 | Improved YOLOv11 algorithm for vehicle damage detectionChangming Zhu, Jie Ji · | YOLOv11, RT-DETR, DETR | ||
| 5 | 2026 | CarDamageEval: Benchmark Evaluation of Car Damage Assessment Using Vision Language ModelsMd Jahid Hasan, Cong Kha Nguyen, Yee Ling Boo, Hamed Jahani · Communications in computer and information science | VLM/MLLM | ||
| 6 | 2025 | Damage Detection Using YOLOv8 AI for Vehicle AssessmentB. Vanathi, R Akshaya, P. Alfina., Gayathri. V. · | YOLOv8 | ||
| 7 | 2025 | Car Damage Detection Based on Multi-View Fusion and Alignment: Dataset and MethodJinbo Peng, Shoubin Dong, Hua Yuan, Xiaorou Zheng · IEEE Transactions on Intelligent Transportation Systems | CDDM | ||
| 8 | 2025 | Vehicle Damage Detection Using Artificial Intelligence: A Systematic Literature ReviewMd Jahid Hasan, Cong Kha Nguyen, Yee Ling Boo, Hamed Jahani · Wiley Interdisciplinary Reviews Data Mining and Knowledge Di | |||
| 9 | 2022 | CarDD: A New Dataset for Vision-based Car Damage DetectionXinkuang Wang, Wenjing Li, Zhongcheng Wu · arXiv | CarDD | ||
| 10 | 2024 | INS-MMBench: A Comprehensive Benchmark for Evaluating LVLMs' Performance in InsuranceChenwei Lin, Hanjia Lyu, Xian Xu, Jiebo Luo · arXiv | GPT-4V/4o, LLaVA, VLM/MLLM | Own/insurer data | |
| 11 | 2025 | Small Dents, Big Impact: A Dataset and Deep Learning Approach for Vehicle Dent DetectionDanish Zia Baig, Mohsin Kamal, Zahid Ullah · arXiv | YOLOv8 | ||
| 12 | 2025 | RF-DETR: Neural Architecture Search for Real-Time Detection TransformersIsaac Robinson, Peter Robicheaux, Matvei Popov, Deva Ramanan · arXiv | DETR, VLM/MLLM, Transfer learning | COCO-car-damage, Roboflow | |
| 13 | 2023 | Detects Damage Car Body using YOLO Deep Learning AlgorithmYonathan Wijaya Gustian, Ben Rahman, Djarot Hindarto, Alessandro Benito Putra Ba · SinkrOn | YOLO | Own/insurer data | precision 70% |
| 14 | 2023 | Deep Learning Based Car Damage Classification and Cost EstimationNamam A. Mohammed, Moayad Yousif Potrus, Abbas M. Ali · ZANCO Journal of Pure and Applied Sciences | Mask R-CNN | Own/insurer data | accuracy 98.5% |
| 15 | 2025 | CAR DAMAGE DETECTION USING DEEP LEARNINGIndukuri, Rahul Varma, Sr. · CSUSB ScholarWorks (California State University, San Bernard | SSD, Inception, MobileNet, Transfer learning | COCO-car-damage | accuracy 94% |
| 16 | 2026 | CropResMoE-50: A Region-Aware Mixture-of-Experts Framework for Fine-Grained Vehicle Damage Detection and Semi-Automated AnnotationKam Ching Cheng, Yong Poh Yu, Tong Ming Lim · Journal of Artificial Intelligence and Technology | ResNet | CarDD, COCO-car-damage | precision 87.7% |
| 17 | 2023 | Assessment of Intelligent Vehicle Damage and Cost Estimator for Insurance CompaniesS. Ramesh Babu B. Naga Rajesh · Journal of Propulsion Technology | VGG, Inception | Own/insurer data | accuracy 94.56% |
| 18 | 2021 | Damaged Vehicle Parts Detection Platforms using Deep Learning TechniqueKundjanasith Thonglek, Norawit Urailertprasert, Patchara Pattiyathanee, Chantana · ECTI Transactions on Computer and Information Technology (EC | Inception | Own/insurer data | accuracy 93.28%, accuracy 98.47% |
| 19 | 2020 | Vehicle-Damage-Detection Segmentation Algorithm Based on Improved Mask RCNNQinghui Zhang, Xianing Chang, Shanfeng Bian · IEEE Access | Mask R-CNN, ResNet, Transfer learning | ||
| 20 | 2025 | Robust anomaly detection through multi-modal autoencoder fusion for small vehicle damage detectionSara Khan, Mehmed Yüksel, Frank Kirchner · Machine Learning with Applications | Ensemble | AUC 92% | |
| 21 | 2025 | Car Damage Detection Analysis using Deep Learning, Computer Vision TechniquesNeeta Verma, P. Santhosh Kumar, Ruchika Yadav, Vidisha Kumar · International Journal of Computer Applications | Transfer learning | accuracy 94% | |
| 22 | 2021 | Automatic Car Damage detection by Hybrid Deep Learning Multi Label ClassificationP. Ebby Darney · Journal of Artificial Intelligence and Capsule Networks | ResNet, Transfer learning | COCO-car-damage | accuracy 90.82% |
| 23 | 2026 | Lightweight Vehicle Damage Detection With NMS‐Free YOLOv10s and AKConvHaichang Jiang, Taozhu Zhou, Yingqian Lao, Chenxin Nie · IET Image Processing | YOLOv10, Xception | mAP 81.48% | |
| 24 | 2026 | An intelligent approach for automated vehicle damage classificationShimaa M. Ouf, Bassant Lotfy, Soha Ahmed · Scientific Reports | YOLOv8, YOLOv5 | precision 87%, precision 90.7% | |
| 25 | 2026 | AI-Based Vehicle Damage and Safety AnalysisM. Karthika M. Vasuki · Zenodo (CERN European Organization for Nuclear Research) | YOLOv10 | mean average precision 88.7% | |
| 26 | 2025 | A System for Automated Vehicle Damage Localization and Severity Estimation Using Deep LearningGeetha M, Chaitra KA · INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING | YOLOv8 | accuracy 91% | |
| 27 | 2025 | CAR Damage Price PredictorG.Uma Maheswari, Jabez Jerin J, John Rathinam, P Sunil · International Research Journal on Advanced Engineering Hub ( | Mask R-CNN, Faster R-CNN, Transfer learning | accuracy 95% | |
| 28 | 2026 | Cross-Architecture Performance Evaluation of Transfer Learning Models for Multi-Class Vehicle Damage Severity ClassificationMochammad Fatih Ulumuddin, Anggay Luri Pramana · SinkrOn | ResNet, VGG, MobileNet, Transfer learning | accuracy 75.76% | |
| 29 | 2023 | Automatic damaged vehicle estimator using enhanced deep learning algorithmJihad Qaddour, Syeda Ayesha Siddiqa · Intelligent Systems with Applications | Mask R-CNN, ResNet, VGG, Inception, Transfer learning | Own/insurer data | |
| 30 | 2024 | Automated Car Damage Assessment Using Computer Vision: Insurance Company Use CaseSergio A. Pérez-Zarate, Daniel Corzo-García, José Luis Pro-Martín, Juan A. Álvar · Applied Sciences | YOLOv8, YOLOv5, Ensemble | Own/insurer data | |
| 31 | 2024 | GroundingCarDD: Text-Guided Multimodal Phrase Grounding for Car Damage DetectionMd Jahid Hasan, Agustinus Nalwan, Kok-Leong Ong, Hamed Jahani · IEEE Access | YOLOv9, DETR | CarDD | |
| 32 | 2025 | On the Application of DiffusionDet to Automatic Car Damage Detection and Classification via High-Performance ComputingVito Arconzo, Gerardo Gorga, Gonzalo Gutierrez, Ahmed Omar · Electronics | Swin | Own/insurer data | |
| 33 | 2024 | Implementation of Car Damage Detection in Insurance ClaimsProf. Soniya Dhotre · INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING | Mask R-CNN | Own/insurer data | |
| 34 | 2025 | Automated Vehicle Damage Detection and Cost Estimator for Insurance CompaniesChinthaprthi Chethan · International Journal for Research in Applied Science and En | YOLOv5 | Own/insurer data | |
| 35 | 2025 | DamageNet: A Dilated Convolution Feature Pyramid Network Mask R‑CNN for Automated Car Damage Detection and SegmentationNazbek Katayev, Zhanna Yessengaliyeva, Zhazira Kozhamkulova, Zhanel Bakirova · International Journal of Advanced Computer Science and Appli | VehiDE | mean Average Precision 85.7%, mIoU 82.3% | |
| 36 | 2023 | Deep Learning-based Damage Detection and Repair Cost Estimation for AutomobilesNarra Suma, N. Savitha · Zenodo (CERN European Organization for Nuclear Research) | MobileNet, Transfer learning | Own/insurer data | |
| 37 | 2024 | Analysis Performance of One-Stage and Two Stage Object Detection Method for Car Damage DetectionHarum Ananda Setyawan, Alhadi Bustamam, Rinaldi Anwar Buyung · International Journal of Advanced Computer Science and Appli | YOLOv8, YOLOv5, ResNet, MobileNet | Own/insurer data | |
| 38 | 2026 | Predicting Vehicle Damage Localization and Severity Estimation using Deep learning.G. Jagan Naik, A. Hemalatha, Dr. A. Vijender · JOURNAL OF ADVANCE AND FUTURE RESEARCH | YOLOv8 | Own/insurer data | |
| 39 | 2026 | AI-Based Vehicle Damage Detection, Cost Estimation, and Insurance Claim Prediction SystemSangeetha Navale, Prachi Raut, Prachi Ukey, Atharva Kulkarni · International Journal of Electrical Electronics and Computer | YOLOv8, Transfer learning | Own/insurer data | |
| 40 | 2026 | DEEP LEARNING BASED VEHICLE DAMAGE ASSESSMENT SYSTEM1Mrs. M. Anusha, 2K. Sravanthi, 3N. Sahithi, 4K. V. S. Janani · International Journal of Data Science and IoT Management Sys | YOLO, Mask R-CNN | Own/insurer data | |
| 41 | 2026 | DEVELOPING A MASK R-CNN MODEL FOR DETECTING CAR DAMAGEMr. MD. IMRAN,T. KEDAR,U. PARDHIV SAI,T. SUPRIYA,P. MANEESH · International Journal of Engineering Research and Science & | Mask R-CNN, Faster R-CNN | COCO-car-damage | |
| 42 | 2020 | Assessing Car Damage using Mask R-CNNSarath P, Soorya M, Shaik Abdul Rahman A, S Suresh Kumar · arXiv | Mask R-CNN | precision 89.5% | |
| 43 | 2024 | Automated vehicle damage classification using the three-quarter view car damage dataset and deep learning approachesDonggeun Lee, Juyeob Lee, Eunil Park · Heliyon | ViT, ResNet, EfficientNet, MobileNet, DenseNet, Ensemble, Transfer learning | ||
| 44 | 2022 | Vehicle Damage Detection and Classification Using Image ProcessingB Mallikarjuna, Arun Kumar K L · International Journal of Advanced Research in Science Commun | YOLO | COCO-car-damage | |
| 45 | 2021 | Deep Learning Based Car Damage Detection, Classification and Severity · International Journal of Advanced Trends in Computer Science | YOLO, ResNet, VGG, DenseNet, Transfer learning | Own/insurer data | |
| 46 | 2022 | A novel data processing approach to detect fraudulent insurance claims for physical damage to carsAhmet YÜCEL · journal of new results in science | Own/insurer data | accuracy 76.56% | |
| 47 | 2026 | From Synthetic Data to Real-World Efficiency: Automated Vehicle Damage Detection and Vehicle Parts SegmentationCátia Loureiro, Sandra Dixe, Cristiana Neto, António Linhares Silva · IEEE Access | YOLOv8, Transfer learning | ||
| 48 | 2022 | VEHICLE DAMAGED DETECTION USING DEEP LEARNING · International Research Journal of Modernization in Engineeri | YOLOv4 | Own/insurer data | |
| 49 | 2025 | HL-YOLO: Improving Vehicle Damage Detection with Heterogeneous Convolutions and Large-Kernel AttentionWeijun Li, Huawei Xie, Peiteng Lin, Liyan Huang · World Electric Vehicle Journal | YOLOv11 | ||
| 50 | 2025 | Enhanced spatial awareness of vision-language model for car damage detectionYunfei Liu, Quan Long · Journal of Electronic Imaging | CLIP, VLM/MLLM, Transfer learning | ||
| 51 | 2024 | CAR DAMAGE DETECTION USING CNNRahul Maruti Dhokane · INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING | Mask R-CNN, VGG | ||
| 52 | 2023 | A Comparative Study of Generative Adversarial Networks for Generating Car Damaged ImagesPhyu Mar Kyu, Kuntpong Woraratpanya · Research Square | ViT, Inception, Transfer learning | ||
| 53 | 2026 | Harris Hawks–tuned severity-aware YOLOv8 instance segmentation framework for vehicle damage assessmentSurjeet Dalal, Yogesh Kumar Sharma, Shakti Kundu, Monika Dhananjay Rokade · Scientific Reports | YOLOv8 | ||
| 54 | 2025 | Deep Learning for Car Damage Detection: A Comprehensive ReviewAbhishek Narsayya Bathula, Khan Faiyaz Ahmed, Gaikwad Vaishnavi, Prof. Rashmi Ma · International Journal of Advanced Research in Science Commun | Transfer learning | ||
| 55 | 2026 | AI Powered Car Damage Detection Using Deep LearningShrujana KC and Prof. Thejaswini MN · International Journal of Advanced Research in Science Commun | YOLOv8 | ||
| 56 | 2025 | Interactive Deep Learning System for Automated Car Damage Detection: Multi-Model Evaluation and Interactive Web DeploymentSailaja Madhu, Bharathi Maddikatla, Ranjitha Padakanti, Vineel Sai Kumar Rampall · | YOLOv9 | ||
| 57 | 2026 | Deep Learning Approach for Vehicle Damage Analysis and Repair cost EstimationVelicheti Srinivas, Garaga Bala Venkata Aditya, Tammisetti Shanmukha Aravind, Vi · INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING | YOLOv5, GPT-4V/4o | ||
| 58 | 2023 | Car Dirtiness and Damage Detection For Automatic Service Recommendation Using Machine Learning TechniquesMohammed Abdullah Khan, Gundlapally Siri Reddy, Tarun Ramavath · International Journal of Scientific Research in Science and | YOLOv5, VGG, Transfer learning | ||
| 59 | 2025 | A New Deep Learning Based Damage Classification Pipeline for Car GlassÖmer Küllü, Anıl Çelik, Pınar Ece Aktan Hatipoğlu, Mehmet Kıvılcım Keleş · Journal of Advanced Research in Natural and Applied Sciences | Ensemble | ||
| 60 | 2022 | Damage Detection of An AutomobileR Melvin Raj, R. Jyosthna, Ramya Madhuri N, Akshay Sunny R · International Journal of Engineering Research in Computer Sc | U-Net, Xception | Own/insurer data | |
| 61 | 2026 | Vehicle Damage Detection and Severity EstimationSarabu Joshna Sarabu Joshna · INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING | YOLOv8, Transfer learning | ||
| 62 | 2025 | Vehicle Damage Detection with YOLO and Visual Embedding FusionHyeon-jin Yu, Hee-Cheol Kim · The Journal of the Korean Institute of Information and Commu | YOLOv8, Swin, ResNet | ||
| 63 | 2026 | Smart-Claim: AI-Based Vehicle Damage Detection and Estimation Using YOLO and Large Language ModelsAsan Nainar M, Jeevanandan B K · Iconic Research and Engineering Journals | YOLO | ||
| 64 | 2026 | CarSenseAI: A Deep-Learning Pipeline for Real-Time Vehicle Damage Detection, Severity Assessment, and Repair-Cost EstimationMithun R, Pavithra S, Gururaj Katti, Lucky Mahawar · Zenodo (CERN European Organization for Nuclear Research) | YOLOv8, Mask R-CNN | ||
| 65 | 2024 | Improved Accuracy of Vehicle Part Detection and Damage Classification using YOLO AlgorithmMan Zhang, Seung-Soo Shin · Journal of Digital Contents Society | YOLO | ||
| 66 | 2026 | COMPARE ONE-STAGE AND TWO-STAGE SEGMENTATION FOR DETECTING VEHICLE DAMAGE IN AN IMAGENataliya I. Boyko, Yurii В. Bodnar · Collection of Scientific Publications NUS | Ensemble | ||
| 67 | 2024 | Car Damage Detection and Patch-to-Patch Self-supervised Image AlignmentHanxiao Chen · arXiv (Cornell University) | Mask R-CNN | ||
| 68 | 2025 | Vehicle Intelligent Damage Assessment Based on Improved YOLOv8-seg飞扬 王 · Modeling and Simulation | YOLOv8 | ||
| 69 | 2026 | Comparative Analysis of CNN Architectures for Vehicle Accident Damage Classification in Intelligent Transportation Systems (ITS)Eren Dağlı, Yavuz Selim Taşpınar, Metin Mutlu Aydın, Rıdvan Ertuğrul Yildirim · Applied Sciences | ResNet | ||
| 70 | 2025 | Research on Damage Detection and Recognition System for Automotive Components Based on Stereo Vision and Deep LearningYi Zhou, Haozhe Zhao · International Journal of Mechanical and Electrical Engineeri | YOLOv5, Transfer learning | ||
| 71 | 2023 | MARS: Mask Attention Refinement with Sequential Quadtree Nodes for Car Damage Instance SegmentationTeerapong Panboonyuen, Naphat Nithisopa, Panin Pienroj, Laphonchai Jirachuphun · arXiv | Mask R-CNN | ||
| 72 | 2022 | Convolutional Neural Networks for vehicle damage detectionRobin Enrico van Ruitenbeek, Sandjai Bhulai · Machine Learning with Applications | Transfer learning | ||
| 73 | 2023 | Vehicle Damage Severity Estimation for Insurance Operations Using In-The-Wild Mobile ImagesDimitrios Mallios, Xiaofei Li, Niall McLaughlin, Jesús Martínez del Rincón · IEEE Access | Own/insurer data | ||
| 74 | 2024 | Automatic detection of potholes using VGG-16 pre-trained network and Convolutional Neural NetworkSatyabrata Swain, Asis Kumar Tripathy · Heliyon | VGG, Transfer learning | accuracy 97.3% | |
| 75 | 2024 | Powering AI-driven car damage identification based on VeHIDE datasetVan-Dung Hoang, Nhan T. Huynh, Nguyen Tran, Khanh Le · Journal of Information and Telecommunication | VehiDE | ||
| 76 | 2022 | Recognizing the Damaged Surface Parts of Cars in the Real Scene Using a Deep Learning FrameworkMahboub Parhizkar, Majid Amirfakhrian · Mathematical Problems in Engineering | ResNet, Inception | ||
| 77 | 2025 | Predicting car accident severity in Northwest Ethiopia: a machine learning approach leveraging driver, environmental, and road conditionsAbraham Keffale Mengistu, Andualem Enyew Gedefaw, Nebebe Demis Baykemagn, Agmasi · Scientific Reports | Ensemble | accuracy 82% | |
| 78 | 2022 | Deep Convolutional Neural Networks With Transfer Learning for Automobile Damage Image ClassificationXiaoguang Tian, Henry Han · Journal of Database Management | Transfer learning | ||
| 79 | 2024 | CES: Cost Estimation System for Enhancing the Processing of Car Insurance ClaimsAhmed Shawky Elbhrawy, Mohamed Belal, Mohamed Sameh Hassanein · Journal of Computing and Communication | YOLO | precision 78.5%, recall 70.24% | |
| 80 | 2024 | An Insurtech Platform to Support Claim Management Through the Automatic Detection and Estimation of Car Damage from PicturesMohab M. H. Atanasious, Valentina Becchetti, Alessandro Giuseppi, Antonio Pietra · Electronics | Own/insurer data | ||
| 81 | 2019 | Preventing Car Damage using CNN and Computer VisionDr.Avinash Sharma, Aaditi Verma, Dhananjay Gupta · International Journal of Innovative Technology and Exploring | Transfer learning | ||
| 82 | 2024 | Enhancing car damage repair cost prediction: Integrating ontology reasoning with regression modelsHamid Ahaggach, Lylia Abrouk A, Eric Lebon · Intelligent Systems with Applications | Own/insurer data | ||
| 83 | 2021 | Modeling and Application of Neural Networks for Automotive Damage AppraisalsFred Poon, Yang Zhang, Jonathon Roach, David Josephs · SMU Scholar (Southern Methodist University) | Mask R-CNN | ||
| 84 | 2023 | Vehicle Insurance Damage DetectionA Jaya Sree, G Aravind, H Lalith, M Yuvraj · International Journal for Research in Applied Science and En | Own/insurer data | ||
| 85 | 2025 | Implementing Redefining Vehicle Damage Detection with Deep Learning Revolutionizing Auto Assessments Using Convolutional Neural NetworkMd Moshiur Rahman, Arif Khan · Journal of Scientific and Technological Research | Own/insurer data | ||
| 86 | 2020 | Car Exterior Damage Detection Using Mask R-CNNPradhip Kumar.R, Sanjeev Patel, S. Shankar · International journal of advance research and innovative ide | Mask R-CNN, Faster R-CNN, Cascade R-CNN, Transfer learning | ||
| 87 | 2026 | Vehicle damage detection with transformer-based multi-object trackingSergio A. Pérez-Zarate, José Luis Pro-Martín, Juan A. Álvarez-García, Miguel Áng · Engineering Applications of Artificial Intelligence | Own/insurer data | ||
| 88 | 2022 | Car Damage Mask RCNNM Dhanushree, K. L. Arunkumar · International Journal of Advanced Research in Science Commun | Mask R-CNN | ||
| 89 | 2024 | INSURANCE AMOUNT PREDICTION BASED ON ACCIDENTAL CAR DAMAGE LEVEL USING CNNB Aditya · INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING | Own/insurer data | ||
| 90 | 2025 | C-DiffDet+: Fusing Global Scene Context with Generative Denoising for High-Fidelity Car Damage DetectionAbdellah Zakaria Sellam, Ilyes Benaissa, Salah Eddine Bekhouche, Abdenour Hadid · arXiv | CarDD | ||
| 91 | 2022 | Automated vehicle inspection model using a deep learning approachMohamed Mostafa Fouad, Karim Malawany, Ahmed Gamil Osman, Hatem Mohamed Amer · Journal of Ambient Intelligence and Humanized Computing | VGG, DenseNet, Ensemble, Transfer learning | accuracy 81%, accuracy 85.5% | |
| 92 | 2023 | Car damage assessment recommendation system using neural networksJason Elroy Martis, M S Sannidhan, C. V. Aravinda, R. Balasubramani · Materials Today Proceedings | |||
| 93 | 2025 | YUNet_LLMClaimReport: An Enhanced Automobile Insurance Fraud Detection and Automated Claim Report Generation Using Large Language ModelsPriyanka D. Kumar, Sountharrajan Sehar · IEEE Access | YOLOv11, U-Net | Kaggle | |
| 94 | 2024 | VERI-D: A new dataset and method for multi-camera vehicle re-identification of damaged cars under varying lighting conditionsLiu Shao, Sos С. Agaian · APL Machine Learning | |||
| 95 | 2024 | K-Nearest Neighbor Algorithm and Case Base Reasoning on Xenia Car Damage Detection Expert SystemBella Ananda, Raissa Amanda Putri · Journal Of Computer Networks Architecture and High Performan | |||
| 96 | 2026 | DeepChiOptimizer for Automated Vehicle Damage Detection Using Hybrid Deep Learning and Feature SelectionÖznur Özaltın · Iranian Journal of Science and Technology Transactions of El | |||
| 97 | 2024 | Vehicle Damage AnalysisShubham Bhoite, Suhas Waghe, Aditya Karbhari, Kaushal Salve · International Journal for Research in Applied Science and En | |||
| 98 | 2026 | Integrating LLMs With Computer Vision: A Geometrically Invariant System for First Notice of Loss in Car InsuranceVito Arconzo, Gerardo Gorga, Gonzalo Gutierrez, Gabriel Pinos · Expert Systems | DETR | CarDD | |
| 99 | 2021 | Development of The Application for Car Audio Parts Detection Damage Using Case Based Reasoning Method and Nearest Neighbor AlgorithmAndika Saputra, Ali Khumaidi · Jurnal Teknik Informatika C I T Medicom | Own/insurer data | ||
| 100 | 2026 | Intact and Damaged Car Parts Detection Using Deep Learning Algorithms Enhanced with Fuzzy LogicSerge Jonathan Kogbeti-Wongo, Usef Faghihi, Fadel Touré, Axel-Frank Tankam N. · SSRN Electronic Journal |
İndirDownload
⬇ Excel (XLSX)100 makale + CarDD sonuçları + bulgular100 papers + CarDD results + findings
⬇ CSVözetlerle birliktewith abstracts⬇ JSONpapers100_db.json
⬇ 100 beceri (MD)100 skills (MD)AI_SKILLS_100_PAPERS.md⬇ CarDD sonuçlarıCarDD resultsresults_cardd.json
⬇ v3 GPU notebookRF-DETR · CarDD · ColabHugging Face ↗datasets/sosa123454321/cekicin-ai-papers