![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F363811%2F0a4f84d4e2b5d34da3904681eb234c94%2Fres-limit.png?generation=1762660421352698&alt=media\")
Using training data for channel II (cyan line) of 104573050-0001.png as an example, which type has the highest potential for highest quality signal reconstruction, one can see that there are multiple areas where signal swings of up to 0.1 mV can not be resolved. See the noisy peaks inside purple ellipses. All pixels in the relevant area have (30, 30, 30) values.
\n
Using training data for channel II (cyan line) of 104573050-0001.png as an example, which type has the highest potential for highest quality signal reconstruction, one can see that there are multiple areas where signal swings of up to 0.1 mV can not be resolved. See the noisy peaks inside purple ellipses. All pixels in the relevant area have (30, 30, 30) values.
\n
Yes, we are limited by both the amplitude and time resolutions, which are functions of the original signal's temporal and amplitude resolutions, the effective image DPI (mv/pixel and seconds/pixel), and the final sampling frequency. Please check out equations A1 and A2 in the Appendix here. A post-extraction low-pass filter can still help. The ripples in your reconstructed signal suggest that the post-extraction resampling filter might not be suitable.
", "rawMarkdown": "Yes, we are limited by both the amplitude and time resolutions, which are functions of the original signal's temporal and amplitude resolutions, the effective image DPI (mv/pixel and seconds/pixel), and the final sampling frequency. Please check out equations A1 and A2 in the Appendix [here](https://doi.org/10.48550/arXiv.2409.16612). A post-extraction low-pass filter can still help. The ripples in your reconstructed signal suggest that the post-extraction resampling filter might not be suitable.", "votes": 1, "replies": [ { "id": 3314066, "postDate": "2025-11-10T04:09:26.917Z", "content": "The ripples are in the training data. The point of my post is that these ripples can't be reliably reconstructed from provided images. I edited my post to make it more clear.
", "rawMarkdown": "The ripples are in the training data. The point of my post is that these ripples can't be reliably reconstructed from provided images. I edited my post to make it more clear.", "votes": 1, "replies": [ { "id": 3314507, "postDate": "2025-11-10T08:04:45.650Z", "content": "you probably cannot judge by visual inspection alone. You need to work out the SNR or MSE values\nYou did not consider subpixel (smudging of neighbours). SNR is capped at 28db by kaggle metric. \nI think if you convert from csv values to plot and reconstruct the plot back to predicteded values, you can get greater than 28db from truth and predicteded values easily for 0001 clean image. You may not able to reconstruct small \"ripples\", but they may not contribute much to the SNR. note that the length of signal in csv file is \"only\" about 2.6x the pixel length in image, which is not a very bad sampling frequency.
\nBigger problem will be the non-0001 images which has distort noise and pixel noise.
", "rawMarkdown": "you probably cannot judge by visual inspection alone. You need to work out the SNR or MSE values\nYou did not consider subpixel (smudging of neighbours). SNR is capped at 28db by kaggle metric. \nI think if you convert from csv values to plot and reconstruct the plot back to predicteded values, you can get greater than 28db from truth and predicteded values easily for 0001 clean image. You may not able to reconstruct small \"ripples\", but they may not contribute much to the SNR. note that the length of signal in csv file is \"only\" about 2.6x the pixel length in image, which is not a very bad sampling frequency.\n\nBigger problem will be the non-0001 images which has distort noise and pixel noise.", "votes": 1 }, { "id": 3314815, "postDate": "2025-11-10T11:20:27.547Z", "content": "I'm curious how you arrived at 28dB limit. I think the limit is at ~25.8dB, see https://www.kaggle.com/competitions/physionet-ecg-image-digitization/discussion/612891#3311018
", "rawMarkdown": "I'm curious how you arrived at 28dB limit. I think the limit is at ~25.8dB, see https://www.kaggle.com/competitions/physionet-ecg-image-digitization/discussion/612891#3311018", "votes": 1 } ] } ] }, { "id": 3313760, "postDate": "2025-11-09T19:00:19.413Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3313588, "author_name": "Reza Sameni", "author_url": "", "post_date": "2025-11-09T14:04:38.890000", "content": "Yes, we are limited by both the amplitude and time resolutions, which are functions of the original signal's temporal and amplitude resolutions, the effective image DPI (mv/pixel and seconds/pixel), and the final sampling frequency. Please check out equations A1 and A2 in the Appendix here. A post-extraction low-pass filter can still help. The ripples in your reconstructed signal suggest that the post-extraction resampling filter might not be suitable.
", "votes": 1, "replies": [ { "id": 3314066, "author_name": "Paul Jurczak", "author_url": "", "post_date": "2025-11-10T04:09:26.917000", "content": "The ripples are in the training data. The point of my post is that these ripples can't be reliably reconstructed from provided images. I edited my post to make it more clear.
", "votes": 1, "replies": [ { "id": 3314507, "author_name": "hengck23", "author_url": "", "post_date": "2025-11-10T08:04:45.650000", "content": "you probably cannot judge by visual inspection alone. You need to work out the SNR or MSE values\nYou did not consider subpixel (smudging of neighbours). SNR is capped at 28db by kaggle metric. \nI think if you convert from csv values to plot and reconstruct the plot back to predicteded values, you can get greater than 28db from truth and predicteded values easily for 0001 clean image. You may not able to reconstruct small \"ripples\", but they may not contribute much to the SNR. note that the length of signal in csv file is \"only\" about 2.6x the pixel length in image, which is not a very bad sampling frequency.
\nBigger problem will be the non-0001 images which has distort noise and pixel noise.
", "votes": 1, "replies": [] }, { "id": 3314815, "author_name": "Paul Jurczak", "author_url": "", "post_date": "2025-11-10T11:20:27.547000", "content": "I'm curious how you arrived at 28dB limit. I think the limit is at ~25.8dB, see https://www.kaggle.com/competitions/physionet-ecg-image-digitization/discussion/612891#3311018
", "votes": 1, "replies": [] } ] } ] }, { "id": 3313760, "author_name": "", "author_url": "", "post_date": "2025-11-09T19:00:19.413000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3313291": "Using training data for channel II (cyan line) of 104573050-0001.png as an example, which type has the highest potential for highest quality signal reconstruction, one can see that there are multiple areas where signal swings of up to 0.1 mV can not be resolved. See the noisy peaks inside purple ellipses. All pixels in the relevant area have (30, 30, 30) values.\n\n", "3313588": "Yes, we are limited by both the amplitude and time resolutions, which are functions of the original signal's temporal and amplitude resolutions, the effective image DPI (mv/pixel and seconds/pixel), and the final sampling frequency. Please check out equations A1 and A2 in the Appendix [here](https://doi.org/10.48550/arXiv.2409.16612). A post-extraction low-pass filter can still help. The ripples in your reconstructed signal suggest that the post-extraction resampling filter might not be suitable.", "3313760": "" } }