I come from a DSP background and have had good luck in the past mixing traditional signal processing with ML. However, in this competition, concatenating \"classical\" image processing features (Gray World, excess red/green, glare correction, etc.) to my DINOv2 embeddings hasn't shown an improvement in CV scores.
\nI suspect the high-dimensional DINO features are \"drowning out\" these few manual signals and/or makes them redundant as the DINO output is likely much more feature-rich. Has anyone successfully used classical image processing features to boost inference in such a complex image-to-value inference problem?
\nFor additional context, my current pipeline extracts the DINOv2 embeddings and provides them to an ensemble of regression models based on recommendations from other discussions and research I have seen. There are likely more effective ways to combine classical features with these embeddings that I have not yet considered as well.
", "messages": [ { "id": "3393725", "postDate": "01/19/2026 16:44:23", "content": "I come from a DSP background and have had good luck in the past mixing traditional signal processing with ML. However, in this competition, concatenating \"classical\" image processing features (Gray World, excess red/green, glare correction, etc.) to my DINOv2 embeddings hasn't shown an improvement in CV scores.
\nI suspect the high-dimensional DINO features are \"drowning out\" these few manual signals and/or makes them redundant as the DINO output is likely much more feature-rich. Has anyone successfully used classical image processing features to boost inference in such a complex image-to-value inference problem?
\nFor additional context, my current pipeline extracts the DINOv2 embeddings and provides them to an ensemble of regression models based on recommendations from other discussions and research I have seen. There are likely more effective ways to combine classical features with these embeddings that I have not yet considered as well.
", "rawMarkdown": "I come from a DSP background and have had good luck in the past mixing traditional signal processing with ML. However, in this competition, concatenating \"classical\" image processing features (Gray World, excess red/green, glare correction, etc.) to my DINOv2 embeddings hasn't shown an improvement in CV scores.\n\nI suspect the high-dimensional DINO features are \"drowning out\" these few manual signals and/or makes them redundant as the DINO output is likely much more feature-rich. Has anyone successfully used classical image processing features to boost inference in such a complex image-to-value inference problem?\n\nFor additional context, my current pipeline extracts the DINOv2 embeddings and provides them to an ensemble of regression models based on recommendations from other discussions and research I have seen. There are likely more effective ways to combine classical features with these embeddings that I have not yet considered as well.", "votes": null }, { "id": "3393820", "postDate": "01/19/2026 22:16:23", "content": "It sounds like right now you are doing feature concatenation. If you want to be overkill you could possibly train an out of fold stacking ensemble. This will weight the DSP and DINOv2 more evenly. This might be a little overkill though. You can also try Gated Fusion which treats your DSP features as data points to use alongside the DINO embeddings. You can take your DSP features and pass them through a small MLP or something and output a vector the same size as your DINOv2 embedding. This can create a gate and suppress DINO features that are less important. This will ultimately help reduce noise before it reaches the final regression.
", "rawMarkdown": "It sounds like right now you are doing feature concatenation. If you want to be overkill you could possibly train an out of fold stacking ensemble. This will weight the DSP and DINOv2 more evenly. This might be a little overkill though. You can also try Gated Fusion which treats your DSP features as data points to use alongside the DINO embeddings. You can take your DSP features and pass them through a small MLP or something and output a vector the same size as your DINOv2 embedding. This can create a gate and suppress DINO features that are less important. This will ultimately help reduce noise before it reaches the final regression.", "votes": null } ], "comments": [ { "id": 3393820, "author_name": "angyscoob", "author_url": "", "post_date": "01/19/2026 22:16:23", "content": "It sounds like right now you are doing feature concatenation. If you want to be overkill you could possibly train an out of fold stacking ensemble. This will weight the DSP and DINOv2 more evenly. This might be a little overkill though. You can also try Gated Fusion which treats your DSP features as data points to use alongside the DINO embeddings. You can take your DSP features and pass them through a small MLP or something and output a vector the same size as your DINOv2 embedding. This can create a gate and suppress DINO features that are less important. This will ultimately help reduce noise before it reaches the final regression.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3393725": "I come from a DSP background and have had good luck in the past mixing traditional signal processing with ML. However, in this competition, concatenating \"classical\" image processing features (Gray World, excess red/green, glare correction, etc.) to my DINOv2 embeddings hasn't shown an improvement in CV scores.\n\nI suspect the high-dimensional DINO features are \"drowning out\" these few manual signals and/or makes them redundant as the DINO output is likely much more feature-rich. Has anyone successfully used classical image processing features to boost inference in such a complex image-to-value inference problem?\n\nFor additional context, my current pipeline extracts the DINOv2 embeddings and provides them to an ensemble of regression models based on recommendations from other discussions and research I have seen. There are likely more effective ways to combine classical features with these embeddings that I have not yet considered as well.", "3393820": "It sounds like right now you are doing feature concatenation. If you want to be overkill you could possibly train an out of fold stacking ensemble. This will weight the DSP and DINOv2 more evenly. This might be a little overkill though. You can also try Gated Fusion which treats your DSP features as data points to use alongside the DINO embeddings. You can take your DSP features and pass them through a small MLP or something and output a vector the same size as your DINOv2 embedding. This can create a gate and suppress DINO features that are less important. This will ultimately help reduce noise before it reaches the final regression." }, "source": "meta" }