Mined pairs: what weak supervision looks like¶
What to look for: before fine-tuning, positives and hard negatives often score alike. That overlap is the only thing the training objective has to remove.
WildMatch never sees a keypoint label. Its training signal is a set of image pairs mined once with the pretrained matcher on the training split: for each anchor photo, the five same-individual photos it scores highest become positives and the five other-individual photos it scores highest become hard negatives. Fine-tuning then pushes the anchor's score with every positive above its score with every negative. Pick an anchor to see its mined pools with the pretrained scores; click a photo to see the correspondences behind that score.
- Scores are the pretrained matcher's image scores on background-removed inputs, the same quantity the method ranks with. Positives (blue) and hard negatives (red) often score alike at this stage: that overlap is what fine-tuning removes.
- Hard negatives come from other individuals the pretrained matcher already finds similar, frequently from the same camera site.
- Anchors shown here were chosen by appearance (daylight colour photos, one per individual); the pools and scores are exactly those in the training index.
Photographs from CzechLynx (Picek et al.), time-closed training split.