![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F0b8652eaf460fa09483d6a48b671a619%2FSelection_989.png?generation=1763264638240666&alt=media\")
to store a large target mask, which is mostly zeros and less than <1% non zeros, in memory, ask chatgpt or gemini for code
\n… to be updated …
\n… current rectification pipleline has 1 to 2 pixel error. i ask chatgpt and he says he can get subpixel accuracy (using a bunch of huerstics like those in corner/line detection for camera calibration/strero vision in computer vision) .. i have yet to try those …
\n… it is probably difficult to separate occlusion of lead waves using segmentation or detect those missing due to dirt patch etc. But it is easy if you use the periodic nature of ECG wave … this is the same as detecting spikes etc. it is like grammar correction in LLM … maybe a denoising or reconstruction transformer head would work?
", "messages": [ { "id": 3328987, "postDate": "2025-11-16T03:45:12.837Z", "content": "
to store a large target mask, which is mostly zeros and less than <1% non zeros, in memory, ask chatgpt or gemini for code
\n… to be updated …
\n… current rectification pipleline has 1 to 2 pixel error. i ask chatgpt and he says he can get subpixel accuracy (using a bunch of huerstics like those in corner/line detection for camera calibration/strero vision in computer vision) .. i have yet to try those …
\n… it is probably difficult to separate occlusion of lead waves using segmentation or detect those missing due to dirt patch etc. But it is easy if you use the periodic nature of ECG wave … this is the same as detecting spikes etc. it is like grammar correction in LLM … maybe a denoising or reconstruction transformer head would work?
", "rawMarkdown": "##Hint 1: increase resolution\n\n\n\nto store a large target mask, which is mostly zeros and less than <1% non zeros, in memory, ask chatgpt or gemini for code\n\n\n##Hint2: refine target mask \n... to be updated ...\n\n\n##Hint3: refine input (improve rectification by post processing)\n... current rectification pipleline has 1 to 2 pixel error. i ask chatgpt and he says he can get subpixel accuracy (using a bunch of huerstics like those in corner/line detection for camera calibration/strero vision in computer vision) .. i have yet to try those ...\n\n\n##Hint4: signal prediction\n... it is probably difficult to separate occlusion of lead waves using segmentation or detect those missing due to dirt patch etc. But it is easy if you use the periodic nature of ECG wave ... this is the same as detecting spikes etc. it is like grammar correction in LLM ... maybe a denoising or reconstruction transformer head would work?", "votes": 39 }, { "id": 3341283, "postDate": "2025-11-20T05:02:50.477Z", "content": "gemini3 has recevied quite a number of positive reviews. So i try it out. It seems that it knows my secret sauce
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd34238b2b5deabb25a6a67c8e0b30ff9%2FSelection_1095.png?generation=1763614945538997&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd9b52a561b2d2ec48e3c54071abfb76b%2FSelection_1094.png?generation=1763614955174218&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F741110cdecbbdefedcf2abb496e42b31%2FSelection_1098.png?generation=1763615575943242&alt=media\")
this works!!
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F0cbd425759fae361605511d1d96915ee%2FSelection_1022.png?generation=1763438380077589&alt=media\")
I guess you are about to break through 20 SNR with this method and increase resolution? By the way, Hint 4 is useful. I improved nearly 3 locally, but only 0.5 on the LB.
", "rawMarkdown": "I guess you are about to break through 20 SNR with this method and increase resolution? By the way, Hint 4 is useful. I improved nearly 3 locally, but only 0.5 on the LB.", "replies": [ { "id": 3335586, "postDate": "2025-11-18T05:23:10.523Z", "content": "Resolution is limited by alignment. if we have good alignment, deep learning or heuristics (which may be easier), we can go beyond 20. my current score is based on resolution x2 only. Simulation on <1 pixel alignment error is 21 snr
", "rawMarkdown": "Resolution is limited by alignment. if we have good alignment, deep learning or heuristics (which may be easier), we can go beyond 20. my current score is based on resolution x2 only. Simulation on <1 pixel alignment error is 21 snr" }, { "id": 3335595, "postDate": "2025-11-18T05:30:29.743Z", "content": "my current rectification has at most 5 pix. detect all grid cell containing signal. just move their corners to the nearest on a reference grid. (need not estimate flow for whole image) just need to rectify cells that contain signal wave or use barycentric coordinates for quadraples
", "rawMarkdown": "my current rectification has at most 5 pix. detect all grid cell containing signal. just move their corners to the nearest on a reference grid. (need not estimate flow for whole image) just need to rectify cells that contain signal wave or use barycentric coordinates for quadraples " }, { "id": 3336449, "postDate": "2025-11-18T13:45:02.807Z", "content": "Predicting the signal (forwards and backwards) through an artifact is a useful way to improve the extract, but can also cause errors if the artifact happens to begin at the start of a sudden change in the waveform. Another way to do this is to try to reconstruct the signal from the other signals that are simultaneously recorded (if the artifact is on the image - if it's on the original recording, the reconstruction might not be as useful). Combining methods and seeing when they agree can help.
", "rawMarkdown": "Predicting the signal (forwards and backwards) through an artifact is a useful way to improve the extract, but can also cause errors if the artifact happens to begin at the start of a sudden change in the waveform. Another way to do this is to try to reconstruct the signal from the other signals that are simultaneously recorded (if the artifact is on the image - if it's on the original recording, the reconstruction might not be as useful). Combining methods and seeing when they agree can help." } ] }, { "id": 3361911, "postDate": "2025-12-04T00:42:54.643Z", "content": "Interesting approach. What does the \"flow\" in this case mean exactly?
", "rawMarkdown": "Interesting approach. What does the \"flow\" in this case mean exactly?" } ] }, { "id": 3331723, "postDate": "2025-11-16T14:04:14.287Z", "content": "this is what subpixel prediction is about
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb51579636d6026e5f0782f50598881d2%2FSelection_997.png?generation=1763301852258785&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F57f1d4e1a74fc948e3d22292c1cc01df%2FSelection_1006.png?generation=1763306580246958&alt=media\")
if i want to predict the warping, i need to predict the line pixels as well … oh i have been doing line prediction for the 3 stages as a by-product
", "rawMarkdown": "this is what subpixel prediction is about\n\n\n\n\n\nif i want to predict the warping, i need to predict the line pixels as well ... oh i have been doing line prediction for the 3 stages as a by-product" }, { "id": 3330252, "postDate": "2025-11-16T08:06:01.637Z", "content": "I more or less confirm the public set are easier images than private set. So if u you adjust huerstics to overfit the public scores … Good luck! make sure your methods are robust to unexpected noise
", "rawMarkdown": "I more or less confirm the public set are easier images than private set. So if u you adjust huerstics to overfit the public scores ... Good luck! make sure your methods are robust to unexpected noise" }, { "id": 3329005, "postDate": "2025-11-16T03:53:59.690Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F1beecf6dc1f19a43755a9012c064f240%2FSelection_993.png?generation=1763265116667296&alt=media\")
current mask (i called it 0001 mask) is create from csv file for reference image in rectification.
\nyou can see 0001 benefits most from this (especially in higher resolution).
\nfor the other type 0003, 0005 … becuase of rectification error, the gain is lower.
\nhence my next step is to make a deep/shallow net to make refined and better traget mask by nudging 0001 mask.
i ask chatgpt to do this and it works …
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F1433010cf171dc6b80f61a33ca3e74ce%2FSelection_994.png?generation=1763265403626547&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fff468aeab8345732133b09737475594f%2FSelection_995.png?generation=1763265422007105&alt=media\")
more on this later ….
", "rawMarkdown": "i ask chatgpt to do this and it works ...\n\n\n\n\nmore on this later ...." } ] }, { "id": 3341892, "postDate": "2025-11-20T13:50:20.820Z", "content": "Simply fix not detecting the peak by line tracing
", "rawMarkdown": "Simply fix not detecting the peak by line tracing", "votes": -1 }, { "id": 3335610, "postDate": "2025-11-18T05:42:29.750Z", "content": "error rate\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F04724ba475e1560dbe1766b5c0c63c40%2FSelection_1023.png?generation=1763444545745269&alt=media\")
The sharp insights and wealth of knowledge of a brilliant genius. The greatest beauty lies in the good intentions of sharing it with others. Those who upvoted this post also showed good intentions (myself included) in sharing their thoughts on the great article. Personally, I was about to give up, thinking it was difficult to make progress, but I gained inspiration beyond the idea.
", "rawMarkdown": "The sharp insights and wealth of knowledge of a brilliant genius. The greatest beauty lies in the good intentions of sharing it with others. Those who upvoted this post also showed good intentions (myself included) in sharing their thoughts on the great article. Personally, I was about to give up, thinking it was difficult to make progress, but I gained inspiration beyond the idea." }, { "id": 3339271, "postDate": "2025-11-19T02:09:25.030Z", "content": "This thread has been incredibly insightful thank you for openly sharing these hints!\nA few key takeaways for me are how much SNR improvement actually depends on early-stage alignment, mask refinement, and leveraging ECG periodicity for reconstruction.
", "rawMarkdown": "This thread has been incredibly insightful thank you for openly sharing these hints!\nA few key takeaways for me are how much SNR improvement actually depends on early-stage alignment, mask refinement, and leveraging ECG periodicity for reconstruction.", "replies": [ { "id": 3339278, "postDate": "2025-11-19T02:14:40.487Z", "content": "for 0001, you can have perfect rectification.
\nyou can compare snr for perfect rectification and predicted rectification (from deep models).
\nyou can even make a graph of snr verus pix error by injecting noise to perfect rectification.
gemini3 has recevied quite a number of positive reviews. So i try it out. It seems that it knows my secret sauce
\n
this works!!
\n
I guess you are about to break through 20 SNR with this method and increase resolution? By the way, Hint 4 is useful. I improved nearly 3 locally, but only 0.5 on the LB.
", "votes": 0, "replies": [ { "id": 3335586, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-18T05:23:10.523000", "content": "Resolution is limited by alignment. if we have good alignment, deep learning or heuristics (which may be easier), we can go beyond 20. my current score is based on resolution x2 only. Simulation on <1 pixel alignment error is 21 snr
", "votes": 0, "replies": [] }, { "id": 3335595, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-18T05:30:29.743000", "content": "my current rectification has at most 5 pix. detect all grid cell containing signal. just move their corners to the nearest on a reference grid. (need not estimate flow for whole image) just need to rectify cells that contain signal wave or use barycentric coordinates for quadraples
", "votes": 0, "replies": [] }, { "id": 3336449, "author_name": "GDClifford", "author_url": "", "post_date": "2025-11-18T13:45:02.807000", "content": "Predicting the signal (forwards and backwards) through an artifact is a useful way to improve the extract, but can also cause errors if the artifact happens to begin at the start of a sudden change in the waveform. Another way to do this is to try to reconstruct the signal from the other signals that are simultaneously recorded (if the artifact is on the image - if it's on the original recording, the reconstruction might not be as useful). Combining methods and seeing when they agree can help.
", "votes": 0, "replies": [] } ] }, { "id": 3361911, "author_name": "Ches Charlemagne", "author_url": "", "post_date": "2025-12-04T00:42:54.643000", "content": "Interesting approach. What does the \"flow\" in this case mean exactly?
", "votes": 0, "replies": [] } ] }, { "id": 3331723, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-16T14:04:14.287000", "content": "this is what subpixel prediction is about
\n
if i want to predict the warping, i need to predict the line pixels as well … oh i have been doing line prediction for the 3 stages as a by-product
", "votes": 0, "replies": [] }, { "id": 3330252, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-16T08:06:01.637000", "content": "I more or less confirm the public set are easier images than private set. So if u you adjust huerstics to overfit the public scores … Good luck! make sure your methods are robust to unexpected noise
", "votes": 0, "replies": [] }, { "id": 3329005, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-16T03:53:59.690000", "content": "
current mask (i called it 0001 mask) is create from csv file for reference image in rectification.
\nyou can see 0001 benefits most from this (especially in higher resolution).
\nfor the other type 0003, 0005 … becuase of rectification error, the gain is lower.
\nhence my next step is to make a deep/shallow net to make refined and better traget mask by nudging 0001 mask.
i ask chatgpt to do this and it works …
\n\n
more on this later ….
", "votes": 0, "replies": [] } ] }, { "id": 3341892, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-20T13:50:20.820000", "content": "Simply fix not detecting the peak by line tracing
", "votes": -1, "replies": [] }, { "id": 3335610, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-18T05:42:29.750000", "content": "error rate\n
The sharp insights and wealth of knowledge of a brilliant genius. The greatest beauty lies in the good intentions of sharing it with others. Those who upvoted this post also showed good intentions (myself included) in sharing their thoughts on the great article. Personally, I was about to give up, thinking it was difficult to make progress, but I gained inspiration beyond the idea.
", "votes": 0, "replies": [] }, { "id": 3339271, "author_name": "Ladipo Samson", "author_url": "", "post_date": "2025-11-19T02:09:25.030000", "content": "This thread has been incredibly insightful thank you for openly sharing these hints!\nA few key takeaways for me are how much SNR improvement actually depends on early-stage alignment, mask refinement, and leveraging ECG periodicity for reconstruction.
", "votes": 0, "replies": [ { "id": 3339278, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-19T02:14:40.487000", "content": "for 0001, you can have perfect rectification.
\nyou can compare snr for perfect rectification and predicted rectification (from deep models).
\nyou can even make a graph of snr verus pix error by injecting noise to perfect rectification.