{
  "id": 664949,
  "title": "Exploring Super Pixel based Region Tokens for Biomass Prediction",
  "url": "/competitions/csiro-biomass/discussion/664949",
  "author_name": "Rajat Nandkumar Shedshyal",
  "post_date": "2025-12-29T06:08:12.744000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I was exploring in this direction, we I would use <strong>DINO</strong> to generate features and then somehow aggregate these features for \"Dry Clover Biomass\", Instead of simple and fixed patch pooling, I was thinking we can go through Super-pixel based approach (like <strong>SLIC</strong>). where DINO feature maps are pooled into regions and tokens are generated that resonate with clovers. </p>\n<p>The Intuition is that biomass depends upon area, density and grouping especially in clover regions as u don't find one are two clovers but a group of them. Pixel level or fixed patches wont help much.</p>\n<p>These tokens are then passed into a Transformer encoder for global reasoning before regression.</p>\n<p>Curious if anyone has tried SuperPixel Pooling? or Region Tokens? or any other similar aggregation techniques in this competition?. Would love to hear your thoughts and give me a direction to procced or any other help is appreciated</p>",
  "messages": [
    {
      "id": 3383014,
      "postDate": "2025-12-29T06:08:12.743Z",
      "content": "<p>I was exploring in this direction, we I would use <strong>DINO</strong> to generate features and then somehow aggregate these features for \"Dry Clover Biomass\", Instead of simple and fixed patch pooling, I was thinking we can go through Super-pixel based approach (like <strong>SLIC</strong>). where DINO feature maps are pooled into regions and tokens are generated that resonate with clovers. </p>\n<p>The Intuition is that biomass depends upon area, density and grouping especially in clover regions as u don't find one are two clovers but a group of them. Pixel level or fixed patches wont help much.</p>\n<p>These tokens are then passed into a Transformer encoder for global reasoning before regression.</p>\n<p>Curious if anyone has tried SuperPixel Pooling? or Region Tokens? or any other similar aggregation techniques in this competition?. Would love to hear your thoughts and give me a direction to procced or any other help is appreciated</p>",
      "rawMarkdown": "I was exploring in this direction, we I would use **DINO** to generate features and then somehow aggregate these features for \"Dry Clover Biomass\", Instead of simple and fixed patch pooling, I was thinking we can go through Super-pixel based approach (like **SLIC**). where DINO feature maps are pooled into regions and tokens are generated that resonate with clovers. \n\nThe Intuition is that biomass depends upon area, density and grouping especially in clover regions as u don't find one are two clovers but a group of them. Pixel level or fixed patches wont help much.\n\nThese tokens are then passed into a Transformer encoder for global reasoning before regression.\n\nCurious if anyone has tried SuperPixel Pooling? or Region Tokens? or any other similar aggregation techniques in this competition?. Would love to hear your thoughts and give me a direction to procced or any other help is appreciated",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3383014": "I was exploring in this direction, we I would use **DINO** to generate features and then somehow aggregate these features for \"Dry Clover Biomass\", Instead of simple and fixed patch pooling, I was thinking we can go through Super-pixel based approach (like **SLIC**). where DINO feature maps are pooled into regions and tokens are generated that resonate with clovers. \n\nThe Intuition is that biomass depends upon area, density and grouping especially in clover regions as u don't find one are two clovers but a group of them. Pixel level or fixed patches wont help much.\n\nThese tokens are then passed into a Transformer encoder for global reasoning before regression.\n\nCurious if anyone has tried SuperPixel Pooling? or Region Tokens? or any other similar aggregation techniques in this competition?. Would love to hear your thoughts and give me a direction to procced or any other help is appreciated"
  }
}