{
  "id": 669620,
  "title": "🥉PhysioNet 98th place solution Training+Inference",
  "url": "/competitions/physionet-ecg-image-digitization/discussion/669620",
  "author_name": "Kirderf",
  "post_date": "2026-01-23T11:05:40.718000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>TRANING + INFERENCE CODE --&gt; <a href=\"https://www.kaggle.com/code/kirderf/physionet-98th-place-solution-training-inference\" target=\"_blank\">https://www.kaggle.com/code/kirderf/physionet-98th-place-solution-training-inference</a></p>\n<p>End-to-end summary (final 98th place solution)\nPipeline overview (training → inference):</p>\n<p>Train source-suffix classifier on training images to predict the acquisition source (0001–0012).\nInference on test images: optional source‑aware preprocessing → Stage 0 (alignment) → Stage 1 (grid rectification) → Stage 2 (Net3 signal extraction) → submission CSV.</p>\n<p>Key results: Public 18.44 / Private 18.33 (98th place).</p>\n<p>Why it works (high level):</p>\n<p>Stage 0/1 explicitly correct camera geometry and grid alignment, which is the dominant error source in this competition.\nStage 2 (Net3) focuses on waveform extraction once the geometry is consistent.\nSource‑aware preprocessing reduces grid/lead distortion on out‑of‑distribution sources, improving SNR and lead consistency.</p>\n<p>Why eca_nfnet_l0 works well here:</p>\n<p>Strong texture + edge sensitivity for ECG grids and lead patterns, with stable feature scales (NFNet design).\nECA attention improves channel selection without heavy overhead, which helps distinguish subtle source‑style differences.\nPerforms well at small resolution (256) with high class imbalance (source classes), making it efficient and accurate for the classifier.\nData flow (simplified): test.png → source classifier → preprocess_by_source → stage0 → stage1 → stage2 → series → submission.csv</p>",
  "messages": [
    {
      "id": 3395637,
      "postDate": "2026-01-23T11:05:40.720Z",
      "content": "<p>TRANING + INFERENCE CODE --&gt; <a href=\"https://www.kaggle.com/code/kirderf/physionet-98th-place-solution-training-inference\" target=\"_blank\">https://www.kaggle.com/code/kirderf/physionet-98th-place-solution-training-inference</a></p>\n<p>End-to-end summary (final 98th place solution)\nPipeline overview (training → inference):</p>\n<p>Train source-suffix classifier on training images to predict the acquisition source (0001–0012).\nInference on test images: optional source‑aware preprocessing → Stage 0 (alignment) → Stage 1 (grid rectification) → Stage 2 (Net3 signal extraction) → submission CSV.</p>\n<p>Key results: Public 18.44 / Private 18.33 (98th place).</p>\n<p>Why it works (high level):</p>\n<p>Stage 0/1 explicitly correct camera geometry and grid alignment, which is the dominant error source in this competition.\nStage 2 (Net3) focuses on waveform extraction once the geometry is consistent.\nSource‑aware preprocessing reduces grid/lead distortion on out‑of‑distribution sources, improving SNR and lead consistency.</p>\n<p>Why eca_nfnet_l0 works well here:</p>\n<p>Strong texture + edge sensitivity for ECG grids and lead patterns, with stable feature scales (NFNet design).\nECA attention improves channel selection without heavy overhead, which helps distinguish subtle source‑style differences.\nPerforms well at small resolution (256) with high class imbalance (source classes), making it efficient and accurate for the classifier.\nData flow (simplified): test.png → source classifier → preprocess_by_source → stage0 → stage1 → stage2 → series → submission.csv</p>",
      "rawMarkdown": "TRANING + INFERENCE CODE --> https://www.kaggle.com/code/kirderf/physionet-98th-place-solution-training-inference\n\nEnd-to-end summary (final 98th place solution)\nPipeline overview (training → inference):\n\nTrain source-suffix classifier on training images to predict the acquisition source (0001–0012).\nInference on test images: optional source‑aware preprocessing → Stage 0 (alignment) → Stage 1 (grid rectification) → Stage 2 (Net3 signal extraction) → submission CSV.\n\nKey results: Public 18.44 / Private 18.33 (98th place).\n\nWhy it works (high level):\n\nStage 0/1 explicitly correct camera geometry and grid alignment, which is the dominant error source in this competition.\nStage 2 (Net3) focuses on waveform extraction once the geometry is consistent.\nSource‑aware preprocessing reduces grid/lead distortion on out‑of‑distribution sources, improving SNR and lead consistency.\n\nWhy eca_nfnet_l0 works well here:\n\nStrong texture + edge sensitivity for ECG grids and lead patterns, with stable feature scales (NFNet design).\nECA attention improves channel selection without heavy overhead, which helps distinguish subtle source‑style differences.\nPerforms well at small resolution (256) with high class imbalance (source classes), making it efficient and accurate for the classifier.\nData flow (simplified): test.png → source classifier → preprocess_by_source → stage0 → stage1 → stage2 → series → submission.csv",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3395637": "TRANING + INFERENCE CODE --> https://www.kaggle.com/code/kirderf/physionet-98th-place-solution-training-inference\n\nEnd-to-end summary (final 98th place solution)\nPipeline overview (training → inference):\n\nTrain source-suffix classifier on training images to predict the acquisition source (0001–0012).\nInference on test images: optional source‑aware preprocessing → Stage 0 (alignment) → Stage 1 (grid rectification) → Stage 2 (Net3 signal extraction) → submission CSV.\n\nKey results: Public 18.44 / Private 18.33 (98th place).\n\nWhy it works (high level):\n\nStage 0/1 explicitly correct camera geometry and grid alignment, which is the dominant error source in this competition.\nStage 2 (Net3) focuses on waveform extraction once the geometry is consistent.\nSource‑aware preprocessing reduces grid/lead distortion on out‑of‑distribution sources, improving SNR and lead consistency.\n\nWhy eca_nfnet_l0 works well here:\n\nStrong texture + edge sensitivity for ECG grids and lead patterns, with stable feature scales (NFNet design).\nECA attention improves channel selection without heavy overhead, which helps distinguish subtle source‑style differences.\nPerforms well at small resolution (256) with high class imbalance (source classes), making it efficient and accurate for the classifier.\nData flow (simplified): test.png → source classifier → preprocess_by_source → stage0 → stage1 → stage2 → series → submission.csv"
  }
}