{ "id": 417001, "title": "28th Solution - KeyNetAffNetHardNet + AdaLAM", "url": "/competitions/image-matching-challenge-2023/discussion/417001", "author_name": "Fang", "post_date": "2023-06-13T19:39:54.236000", "votes": 17, "comment_count": 0, "views": 0, "content": "

Overview:

\n

We utilized a combination of KeyNet detector, AffNet, and HardNet descriptor for feature detection and matching in our approach.

\n

The matching process consisted of two stages:

\n
    \n

  1. In the first stage, we employed AdaLAM to identify initial candidate matches. We found that AdaLAM performs exceptionally well in finding candidate matches compared to the smnn matcher.

    2. \n

  2. Acknowledging that Structure-from-Motion (SfM) tends to favor long feature tracks spanning multiple images, we conducted a second round of local feature matching around each pair of candidate matches identified in Stage 1. Specifically, we searched for k=10 nearest neighbors around each candidate match pair and utilized AdaLAM to further expand the number of matches between the two images. This approach has enhanced the connections of feature tracks among overlapping images. Previously, we observed that although we could find suitable matches in image stereo pairs, these features could not be linked as image tracks, resulting in inferior SfM performance.

    3. \n
\n

Other Tricks:

\n
    \n
  1. Enable OriNet
    2. \n
  2. Increase number of detected features to 8192
    3. \n
  3. Resize the longer edge of image to 1600
    4. \n
\n

Not Work:

\n
    \n
  1. Grouping detected features in 2x2 pixel grids, which provided the best connection across all images and allowed us to establish tracks over multiple images even with minimal overlap. Unfortunately, the overall score was low. (not sure if we could improve the performance by relocating the grouped feature matches)
    2. \n
  2. Clustering image features using DBSCAN and conducting matching within the bounding box defined around each cluster. Although this approach offered some improvements, the impact was limited.
    3. \n
  3. Concatenating HardNet descriptors generated on images with different scales (Not fully tested)
    4. \n
", "messages": [ { "id": 2301326, "postDate": "2023-06-13T19:39:54.237Z", "content": "

Overview:

\n

We utilized a combination of KeyNet detector, AffNet, and HardNet descriptor for feature detection and matching in our approach.

\n

The matching process consisted of two stages:

\n
    \n

  1. In the first stage, we employed AdaLAM to identify initial candidate matches. We found that AdaLAM performs exceptionally well in finding candidate matches compared to the smnn matcher.

    2. \n

  2. Acknowledging that Structure-from-Motion (SfM) tends to favor long feature tracks spanning multiple images, we conducted a second round of local feature matching around each pair of candidate matches identified in Stage 1. Specifically, we searched for k=10 nearest neighbors around each candidate match pair and utilized AdaLAM to further expand the number of matches between the two images. This approach has enhanced the connections of feature tracks among overlapping images. Previously, we observed that although we could find suitable matches in image stereo pairs, these features could not be linked as image tracks, resulting in inferior SfM performance.

    3. \n
\n

Other Tricks:

\n
    \n
  1. Enable OriNet
    2. \n
  2. Increase number of detected features to 8192
    3. \n
  3. Resize the longer edge of image to 1600
    4. \n
\n

Not Work:

\n
    \n
  1. Grouping detected features in 2x2 pixel grids, which provided the best connection across all images and allowed us to establish tracks over multiple images even with minimal overlap. Unfortunately, the overall score was low. (not sure if we could improve the performance by relocating the grouped feature matches)
    2. \n
  2. Clustering image features using DBSCAN and conducting matching within the bounding box defined around each cluster. Although this approach offered some improvements, the impact was limited.
    3. \n
  3. Concatenating HardNet descriptors generated on images with different scales (Not fully tested)
    4. \n
", "rawMarkdown": "**Overview:** \n \nWe utilized a combination of KeyNet detector, AffNet, and HardNet descriptor for feature detection and matching in our approach.\n\nThe matching process consisted of two stages:\n\n1. In the first stage, we employed AdaLAM to identify initial candidate matches. We found that AdaLAM performs exceptionally well in finding candidate matches compared to the smnn matcher.\n\n2. Acknowledging that Structure-from-Motion (SfM) tends to favor long feature tracks spanning multiple images, we conducted a second round of local feature matching around each pair of candidate matches identified in Stage 1. Specifically, we searched for k=10 nearest neighbors around each candidate match pair and utilized AdaLAM to further expand the number of matches between the two images. This approach has enhanced the connections of feature tracks among overlapping images. Previously, we observed that although we could find suitable matches in image stereo pairs, these features could not be linked as image tracks, resulting in inferior SfM performance.\n\n**Other Tricks:** \n\n1. Enable OriNet\n2. Increase number of detected features to 8192\n3. Resize the longer edge of image to 1600\n\n**Not Work:** \n\n1. Grouping detected features in 2x2 pixel grids, which provided the best connection across all images and allowed us to establish tracks over multiple images even with minimal overlap. Unfortunately, the overall score was low. (not sure if we could improve the performance by relocating the grouped feature matches)\n2. Clustering image features using DBSCAN and conducting matching within the bounding box defined around each cluster. Although this approach offered some improvements, the impact was limited.\n3. Concatenating HardNet descriptors generated on images with different scales (Not fully tested)", "votes": 17 } ], "comments": [], "raw_markdown_by_id": { "2301326": "**Overview:** \n \nWe utilized a combination of KeyNet detector, AffNet, and HardNet descriptor for feature detection and matching in our approach.\n\nThe matching process consisted of two stages:\n\n1. In the first stage, we employed AdaLAM to identify initial candidate matches. We found that AdaLAM performs exceptionally well in finding candidate matches compared to the smnn matcher.\n\n2. Acknowledging that Structure-from-Motion (SfM) tends to favor long feature tracks spanning multiple images, we conducted a second round of local feature matching around each pair of candidate matches identified in Stage 1. Specifically, we searched for k=10 nearest neighbors around each candidate match pair and utilized AdaLAM to further expand the number of matches between the two images. This approach has enhanced the connections of feature tracks among overlapping images. Previously, we observed that although we could find suitable matches in image stereo pairs, these features could not be linked as image tracks, resulting in inferior SfM performance.\n\n**Other Tricks:** \n\n1. Enable OriNet\n2. Increase number of detected features to 8192\n3. Resize the longer edge of image to 1600\n\n**Not Work:** \n\n1. Grouping detected features in 2x2 pixel grids, which provided the best connection across all images and allowed us to establish tracks over multiple images even with minimal overlap. Unfortunately, the overall score was low. (not sure if we could improve the performance by relocating the grouped feature matches)\n2. Clustering image features using DBSCAN and conducting matching within the bounding box defined around each cluster. Although this approach offered some improvements, the impact was limited.\n3. Concatenating HardNet descriptors generated on images with different scales (Not fully tested)" } }