We are thrilled to win a gold medal in this shake-up competition. Firstly, we'd like to express our gratitude to the organizers for hosting such an great game and for open-sourcing the entire workflow—from EDA to data processing, training, and submission. This significantly lowered the barrier to entry for us. We also want to thank all of the open-sourced authors for their contributions. Finally, a special thanks to my teammate, @forcewithme . We poured much effort into this competition, and we're so happy about the rewarding outcome
\nOverall, our inference pipeline consists of two stages. The first stage uses a very low threshold to recall as many candidates as possible, while the second stage applies an appropriate threshold to make the final decision.
\nRegarding training, we coudn't find a stable way throughout the entire process. In the end, all models we used were either YOLOv8 or YOLO11 models trained simultaneously on both external data @brendanartley and competition data, sharing the same training configuration.
\n| Submission | \nA | \nB | \nC (Selected) | \nD(Selected) | \nE | \n
|---|---|---|---|---|---|
| Overall Scheme | \nStd+S2 | \nParallel[Std+SAHI]+S2 | \nSTD+S2 | \nParallel[Std+SAHI]+S2 | \nSTD+S2 | \n
| Stage 1A Config | \nyolo8l res:960/1280/832 | \nyolov8l no sam yolo11l sam:1/4 res:960 | \nyolov8 no sam res:960 yolov11 sam:1/2 res 960 | \nyolo8l no sam res:960 | \nyolov8l res:960 | \n
| Stage 1B(SAHI) | \nN/A | \nyolo8l sam:1/3 patch:768 | \nN/A | \nyolo11-cz model sam:1/4 patch:640 | \nN/A | \n
| Stage 2 | \nyolo8l z1.5 res512 yolo11-cz z1.5 res640 yolo8l-cz z2 res640 | \nyolo8l z1.5 res512 yolo11-cz z3 res640 yolo8l-cz z2 res640 | \nyolo8l z1.5 res512 yolo11-cz z1.5 res512 yolo8l-cz z2 res640 | \nyolo8l z1.5 res512 | \nyolo8l z1.5 res512 | \n
| Public Score | \n0.837 | \n0.853 | \n0.856 | \n0.856 | \n0.860 | \n
| Private Score | \n0.853 | \n0.855 | \n0.852 | \n0.832 | \n0.826 | \n
Standard: Public inferece pipeline.sahi: Sahi inference pipeline.S2: Stage2 inference. Infernce on zoomin+rotate+crop images, with almost same code as Standard.sam{}: The sample rate applied to the SAHI processingno sam: No sampling, meaning taking all of the slices.max_image_side: Maximum dimension allowed for input images before slicingpatches_per_side: Number of grid divisions (n x n) for each slicepatch_size: Target dimension for each patch (e.g., 768x768 pixels)res: image resolution.-cz z{1.5/2/3}: Means this is a 'local' model, 'cz' means crop and zoomin. And z{} indicates the zoom-in scale while inference. Any questions or suggestions are welcome. Happy kaggling!
", "messages": [ { "id": "3217969", "postDate": "06/05/2025 15:51:07", "content": "We are thrilled to win a gold medal in this shake-up competition. Firstly, we'd like to express our gratitude to the organizers for hosting such an great game and for open-sourcing the entire workflow—from EDA to data processing, training, and submission. This significantly lowered the barrier to entry for us. We also want to thank all of the open-sourced authors for their contributions. Finally, a special thanks to my teammate, @forcewithme . We poured much effort into this competition, and we're so happy about the rewarding outcome
\nOverall, our inference pipeline consists of two stages. The first stage uses a very low threshold to recall as many candidates as possible, while the second stage applies an appropriate threshold to make the final decision.
\nRegarding training, we coudn't find a stable way throughout the entire process. In the end, all models we used were either YOLOv8 or YOLO11 models trained simultaneously on both external data @brendanartley and competition data, sharing the same training configuration.
\n| Submission | \nA | \nB | \nC (Selected) | \nD(Selected) | \nE | \n
|---|---|---|---|---|---|
| Overall Scheme | \nStd+S2 | \nParallel[Std+SAHI]+S2 | \nSTD+S2 | \nParallel[Std+SAHI]+S2 | \nSTD+S2 | \n
| Stage 1A Config | \nyolo8l res:960/1280/832 | \nyolov8l no sam yolo11l sam:1/4 res:960 | \nyolov8 no sam res:960 yolov11 sam:1/2 res 960 | \nyolo8l no sam res:960 | \nyolov8l res:960 | \n
| Stage 1B(SAHI) | \nN/A | \nyolo8l sam:1/3 patch:768 | \nN/A | \nyolo11-cz model sam:1/4 patch:640 | \nN/A | \n
| Stage 2 | \nyolo8l z1.5 res512 yolo11-cz z1.5 res640 yolo8l-cz z2 res640 | \nyolo8l z1.5 res512 yolo11-cz z3 res640 yolo8l-cz z2 res640 | \nyolo8l z1.5 res512 yolo11-cz z1.5 res512 yolo8l-cz z2 res640 | \nyolo8l z1.5 res512 | \nyolo8l z1.5 res512 | \n
| Public Score | \n0.837 | \n0.853 | \n0.856 | \n0.856 | \n0.860 | \n
| Private Score | \n0.853 | \n0.855 | \n0.852 | \n0.832 | \n0.826 | \n
Standard: Public inferece pipeline.sahi: Sahi inference pipeline.S2: Stage2 inference. Infernce on zoomin+rotate+crop images, with almost same code as Standard.sam{}: The sample rate applied to the SAHI processingno sam: No sampling, meaning taking all of the slices.max_image_side: Maximum dimension allowed for input images before slicingpatches_per_side: Number of grid divisions (n x n) for each slicepatch_size: Target dimension for each patch (e.g., 768x768 pixels)res: image resolution.-cz z{1.5/2/3}: Means this is a 'local' model, 'cz' means crop and zoomin. And z{} indicates the zoom-in scale while inference. Any questions or suggestions are welcome. Happy kaggling!
", "rawMarkdown": "# Overview\nWe are thrilled to win a gold medal in this shake-up competition. Firstly, we'd like to express our gratitude to the organizers for hosting such an great game and for open-sourcing the entire workflow—from EDA to data processing, training, and submission. This significantly lowered the barrier to entry for us. We also want to thank all of the open-sourced authors for their contributions. Finally, a special thanks to my teammate, @forcewithme . We poured much effort into this competition, and we're so happy about the rewarding outcome\n\nOverall, our inference pipeline consists of two stages. The first stage uses a very low threshold to recall as many candidates as possible, while the second stage applies an appropriate threshold to make the final decision.\n\nRegarding training, we coudn't find a stable way throughout the entire process. In the end, all models we used were either YOLOv8 or YOLO11 models trained simultaneously on both external data @brendanartley and competition data, sharing the same training configuration.\n\n## Inference Pipeline\n\n1. **Two-Stage Detection Pipeline:** \n * **Stage 1 (Candidate Generation):** Identifies initial potential motor locations. This stage uses standard model ensembles or SAHI-based ensembles.\n * **Stage 2 (Candidate filter):** Further processes and validates candidates from Stage 1.\n2. **Ensemble Models:** Multiple models or different configurations of the same model are used in combination to enhance detection accuracy and robustness, which applied in both Stage 1 and Stage 2.\n3. **Test-Time Augmentation (TTA):** Since the test dataset has large variance in image size, we use multiple resolutions for inference to capture features at different scales.\n4. **Bypass Logic & Midpoint Reasoning:**\n * **Bypass Logic:** Skip Stage 2 if Stage 1 generates highly confident detections, which improving efficiency with no cost. The bypass threshold is set to 0.6. \n * **Midpoint Reasoning:** If top2 detections are very confident and close(either in Stage 1 or after Stage 2), a new detection point at their geometric midpoint with a slightly boosted confidence will be created returned as prediction point.\n5. **SAHI:** In some slides in Stage 1(if Stage1 doesn't bypass), we use SAHI method to devide large tomogram slices into non-overlapping patches for inference, then merges results to enhance multi-scale detection.\n6. **Rotation-based Refinement with zoom in in stage 2:** Instead of rotation around the candidate, we crop a rotated zoomed-in square around the target, and pad the original image slice with the mean pixel value. \n\n---\n\n## Model training\n\n1. We trained yolov8l or yolov11l with ultralytics.\n2. All of the models are trained with a mixture of external data('trust' 3) and competition data('trust' 4).\n3. The external dataset is fixed by @tatamikenn .\n4. We trained some 'local' model by random cropping around target while training. These models are trained for sahi or stage2.\n\n## Submission Comparison\n\n| Submission | A | B | C (Selected) | D(Selected) |E|\n|:-------------------|:------------|:-------------|:-------------|:-------------|:-------------|\n| **Overall Scheme** | Std+S2 | Parallel[Std+SAHI]+S2 | STD+S2 | Parallel[Std+SAHI]+S2 |STD+S2\n| **Stage 1A Config**| yolo8l@doonut Congratulations on the gold medal, and thanks for the mention!
\nI have a couple of questions:
\nHi @tatamikenn ,
\n\n", "rawMarkdown": "Hi @tatamikenn ,\n\n> Q: How many candidate detections are passed to stage 2?\nA: The number of candidates varies in different submission. But overall it is in range(3,6).>\n\n> Q: How is the \"trust\" level of the dataset utilized in your pipeline?\nA: Do you mean the \"trust\" level in the stage 2 of inference pipeline？ We didn't set a `trust` in inference pipeline.", "votes": null } ], "comments": [ { "id": 3218407, "author_name": "tatamikenn", "author_url": "", "post_date": "06/06/2025 06:33:58", "content": "Q: How many candidate detections are passed to stage 2?
\n
\n A: The number of candidates varies in different submission. But overall it is in range(3,6).>Q: How is the \"trust\" level of the dataset utilized in your pipeline?
\n
\n A: Do you mean the \"trust\" level in the stage 2 of inference pipeline？ We didn't set atrustin inference pipeline.
@doonut Congratulations on the gold medal, and thanks for the mention!
\nI have a couple of questions:
\nHi @tatamikenn ,
\n\n", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3217969": "# Overview\nWe are thrilled to win a gold medal in this shake-up competition. Firstly, we'd like to express our gratitude to the organizers for hosting such an great game and for open-sourcing the entire workflow—from EDA to data processing, training, and submission. This significantly lowered the barrier to entry for us. We also want to thank all of the open-sourced authors for their contributions. Finally, a special thanks to my teammate, @forcewithme . We poured much effort into this competition, and we're so happy about the rewarding outcome\n\nOverall, our inference pipeline consists of two stages. The first stage uses a very low threshold to recall as many candidates as possible, while the second stage applies an appropriate threshold to make the final decision.\n\nRegarding training, we coudn't find a stable way throughout the entire process. In the end, all models we used were either YOLOv8 or YOLO11 models trained simultaneously on both external data @brendanartley and competition data, sharing the same training configuration.\n\n## Inference Pipeline\n\n1. **Two-Stage Detection Pipeline:** \n * **Stage 1 (Candidate Generation):** Identifies initial potential motor locations. This stage uses standard model ensembles or SAHI-based ensembles.\n * **Stage 2 (Candidate filter):** Further processes and validates candidates from Stage 1.\n2. **Ensemble Models:** Multiple models or different configurations of the same model are used in combination to enhance detection accuracy and robustness, which applied in both Stage 1 and Stage 2.\n3. **Test-Time Augmentation (TTA):** Since the test dataset has large variance in image size, we use multiple resolutions for inference to capture features at different scales.\n4. **Bypass Logic & Midpoint Reasoning:**\n * **Bypass Logic:** Skip Stage 2 if Stage 1 generates highly confident detections, which improving efficiency with no cost. The bypass threshold is set to 0.6. \n * **Midpoint Reasoning:** If top2 detections are very confident and close(either in Stage 1 or after Stage 2), a new detection point at their geometric midpoint with a slightly boosted confidence will be created returned as prediction point.\n5. **SAHI:** In some slides in Stage 1(if Stage1 doesn't bypass), we use SAHI method to devide large tomogram slices into non-overlapping patches for inference, then merges results to enhance multi-scale detection.\n6. **Rotation-based Refinement with zoom in in stage 2:** Instead of rotation around the candidate, we crop a rotated zoomed-in square around the target, and pad the original image slice with the mean pixel value. \n\n---\n\n## Model training\n\n1. We trained yolov8l or yolov11l with ultralytics.\n2. All of the models are trained with a mixture of external data('trust' 3) and competition data('trust' 4).\n3. The external dataset is fixed by @tatamikenn .\n4. We trained some 'local' model by random cropping around target while training. These models are trained for sahi or stage2.\n\n## Submission Comparison\n\n| Submission | A | B | C (Selected) | D(Selected) |E|\n|:-------------------|:------------|:-------------|:-------------|:-------------|:-------------|\n| **Overall Scheme** | Std+S2 | Parallel[Std+SAHI]+S2 | STD+S2 | Parallel[Std+SAHI]+S2 |STD+S2\n| **Stage 1A Config**| yolo8lQ: How many candidate detections are passed to stage 2?
\n
\n A: The number of candidates varies in different submission. But overall it is in range(3,6).>Q: How is the \"trust\" level of the dataset utilized in your pipeline?
\n
\n A: Do you mean the \"trust\" level in the stage 2 of inference pipeline？ We didn't set atrustin inference pipeline.