{
  "id": 509825,
  "title": "transparent or symmetry-and-repeats is a key to get high score in LB ? ",
  "url": "/competitions/image-matching-challenge-2024/discussion/509825",
  "author_name": "",
  "post_date": "2024-06-04T01:00:57.354707700Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>This competition gave me a lot of chances to study image processing about 3D, thanks<br>\nI'm wondering about the test data distribution because there is a big difference between local validation and LB(test).</p>\n<p>I used images sampled up to 110 per scene for validation,  and except category of transparent and symmetry-and-repeats, I got big improvement from published notebook.<br>\nHowever, the pipeline does not score well in LB, so I'm wondering that test data have a lot of those category.</p>\n<p>Here is my validation and LB result.  My final pipline is</p>\n<ul>\n<li>rotation matching </li>\n<li>detecting more points when scene inclued transparnt category</li>\n<li>use MAGSAC </li>\n<li>exhaustive matching</li>\n<li>aliked + disk + doghardnet</li>\n</ul>\n<p>The validation result is the mean of 4-times-run of pipeline, which has a purpose to check std, min and max of scores by colmap randomness.</p>\n<table>\n<thead>\n<tr>\n<th>exp</th>\n<th>global</th>\n<th>church</th>\n<th>dioscuri</th>\n<th>lizard</th>\n<th>multi-temporal-temple</th>\n<th>pond</th>\n<th>glass_cup</th>\n<th>glass_cylinder</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>published notebook (LB=0.13)</td>\n<td>0.289</td>\n<td>0.300</td>\n<td>0.164</td>\n<td>0.734</td>\n<td>0.228</td>\n<td>0.553</td>\n<td>0.023</td>\n<td>0.023</td>\n</tr>\n<tr>\n<td>final sub (LB=0.15)</td>\n<td>0.386</td>\n<td>0.301</td>\n<td>0.515</td>\n<td>0.796</td>\n<td>0.430</td>\n<td>0.590</td>\n<td>0.025</td>\n<td>0.045</td>\n</tr>\n</tbody>\n</table>",
  "messages": [
    {
      "id": "2853790",
      "postDate": "06/04/2024 01:00:57",
      "content": "<p>This competition gave me a lot of chances to study image processing about 3D, thanks<br>\nI'm wondering about the test data distribution because there is a big difference between local validation and LB(test).</p>\n<p>I used images sampled up to 110 per scene for validation,  and except category of transparent and symmetry-and-repeats, I got big improvement from published notebook.<br>\nHowever, the pipeline does not score well in LB, so I'm wondering that test data have a lot of those category.</p>\n<p>Here is my validation and LB result.  My final pipline is</p>\n<ul>\n<li>rotation matching </li>\n<li>detecting more points when scene inclued transparnt category</li>\n<li>use MAGSAC </li>\n<li>exhaustive matching</li>\n<li>aliked + disk + doghardnet</li>\n</ul>\n<p>The validation result is the mean of 4-times-run of pipeline, which has a purpose to check std, min and max of scores by colmap randomness.</p>\n<table>\n<thead>\n<tr>\n<th>exp</th>\n<th>global</th>\n<th>church</th>\n<th>dioscuri</th>\n<th>lizard</th>\n<th>multi-temporal-temple</th>\n<th>pond</th>\n<th>glass_cup</th>\n<th>glass_cylinder</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>published notebook (LB=0.13)</td>\n<td>0.289</td>\n<td>0.300</td>\n<td>0.164</td>\n<td>0.734</td>\n<td>0.228</td>\n<td>0.553</td>\n<td>0.023</td>\n<td>0.023</td>\n</tr>\n<tr>\n<td>final sub (LB=0.15)</td>\n<td>0.386</td>\n<td>0.301</td>\n<td>0.515</td>\n<td>0.796</td>\n<td>0.430</td>\n<td>0.590</td>\n<td>0.025</td>\n<td>0.045</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "This competition gave me a lot of chances to study image processing about 3D, thanks\nI'm wondering about the test data distribution because there is a big difference between local validation and LB(test).\n\nI used images sampled up to 110 per scene for validation,  and except category of transparent and symmetry-and-repeats, I got big improvement from published notebook.\nHowever, the pipeline does not score well in LB, so I'm wondering that test data have a lot of those category.\n\nHere is my validation and LB result.  My final pipline is\n- rotation matching \n- detecting more points when scene inclued transparnt category\n- use MAGSAC \n- exhaustive matching\n- aliked + disk + doghardnet\n\n\n\n\nThe validation result is the mean of 4-times-run of pipeline, which has a purpose to check std, min and max of scores by colmap randomness.\n\n| exp |global | church | dioscuri | lizard | multi-temporal-temple | pond | glass_cup | glass_cylinder | \n| --- | --- | --- | --- | --- | --- | --- | --- | --- | \n| published notebook (LB=0.13) | 0.289 | 0.300 | 0.164 | 0.734 | 0.228 | 0.553 | 0.023 | 0.023 | \n| final sub (LB=0.15)| 0.386 | 0.301 | 0.515 | 0.796 | 0.430 | 0.590 | 0.025 | 0.045 |",
      "votes": null
    },
    {
      "id": "2853869",
      "postDate": "06/04/2024 02:30:23",
      "content": "<p>We are preparing our solution write-up, but short answer is yes =) Symmetry is enormously hard, but transparent scenes in this comp are pretty solvable.</p>",
      "rawMarkdown": "We are preparing our solution write-up, but short answer is yes =) Symmetry is enormously hard, but transparent scenes in this comp are pretty solvable.",
      "votes": null
    },
    {
      "id": "2853918",
      "postDate": "06/04/2024 03:20:39",
      "content": "<p>Congratulations on winning the championship🥳! Transparent scenarios are really challenging for me. I have reviewed the literature and feel that this is not an easy problem to solve. I look forward to your brilliant solutions!</p>",
      "rawMarkdown": "Congratulations on winning the championship🥳! Transparent scenarios are really challenging for me. I have reviewed the literature and feel that this is not an easy problem to solve. I look forward to your brilliant solutions!",
      "votes": null
    },
    {
      "id": "2853919",
      "postDate": "06/04/2024 03:23:25",
      "content": "<p>For me，<br>\nchurch 0.19<br>\ndioscuri 0.60<br>\nAfter some attempts, the transparent scenarios were not included in the verification scope.</p>",
      "rawMarkdown": "For me，\nchurch 0.19\ndioscuri 0.60\nAfter some attempts, the transparent scenarios were not included in the verification scope.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2853869,
      "author_name": "vostankovich",
      "author_url": "",
      "post_date": "06/04/2024 02:30:23",
      "content": "<p>We are preparing our solution write-up, but short answer is yes =) Symmetry is enormously hard, but transparent scenes in this comp are pretty solvable.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2853918,
          "author_name": "pamilovedl",
          "author_url": "",
          "post_date": "06/04/2024 03:20:39",
          "content": "<p>Congratulations on winning the championship🥳! Transparent scenarios are really challenging for me. I have reviewed the literature and feel that this is not an easy problem to solve. I look forward to your brilliant solutions!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2853919,
      "author_name": "pamilovedl",
      "author_url": "",
      "post_date": "06/04/2024 03:23:25",
      "content": "<p>For me，<br>\nchurch 0.19<br>\ndioscuri 0.60<br>\nAfter some attempts, the transparent scenarios were not included in the verification scope.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2853790": "This competition gave me a lot of chances to study image processing about 3D, thanks\nI'm wondering about the test data distribution because there is a big difference between local validation and LB(test).\n\nI used images sampled up to 110 per scene for validation,  and except category of transparent and symmetry-and-repeats, I got big improvement from published notebook.\nHowever, the pipeline does not score well in LB, so I'm wondering that test data have a lot of those category.\n\nHere is my validation and LB result.  My final pipline is\n- rotation matching \n- detecting more points when scene inclued transparnt category\n- use MAGSAC \n- exhaustive matching\n- aliked + disk + doghardnet\n\n\n\n\nThe validation result is the mean of 4-times-run of pipeline, which has a purpose to check std, min and max of scores by colmap randomness.\n\n| exp |global | church | dioscuri | lizard | multi-temporal-temple | pond | glass_cup | glass_cylinder | \n| --- | --- | --- | --- | --- | --- | --- | --- | --- | \n| published notebook (LB=0.13) | 0.289 | 0.300 | 0.164 | 0.734 | 0.228 | 0.553 | 0.023 | 0.023 | \n| final sub (LB=0.15)| 0.386 | 0.301 | 0.515 | 0.796 | 0.430 | 0.590 | 0.025 | 0.045 |",
    "2853869": "We are preparing our solution write-up, but short answer is yes =) Symmetry is enormously hard, but transparent scenes in this comp are pretty solvable.",
    "2853918": "Congratulations on winning the championship🥳! Transparent scenarios are really challenging for me. I have reviewed the literature and feel that this is not an easy problem to solve. I look forward to your brilliant solutions!",
    "2853919": "For me，\nchurch 0.19\ndioscuri 0.60\nAfter some attempts, the transparent scenarios were not included in the verification scope."
  },
  "source": "meta"
}