TODO: MaskGeneration

[51N84D]

• [ ] Training on Unity Urban + Suburban Méli
• [ ] Training on WD Méli
• [ ] Training on heavier flooded images (Unity)
• [ ] Conditioning on depth and segmentation Méli
• [ ] Create standard test set (25 images) to evaluate MaskGen on Sun
• [ ] Determine evaluation metrics for MaskGen Sun metrics paper
  • [ ] M-IOU
  • [ ] Pixel Accuracy
  • [ ] Hole detection
• [ ] Domain Adaptation TianYu
  • [ ] Feature discriminator operates on each domain (3 domains -> 3 classes)
  • [ ] Try Sim2Real

[vict0rsch]

(write your names next to items you want to own 🥇)

[melisandeteng]

Depth maps of the WatchDogs dataset computed with MegaDepth model can be found in /network/tmp1/ccai/data/Ubisoft_LaForge_ClimateChange_DataSet/ . They were computed on the non-flooded images resized to 512x384 pixels. Depth can be misestimated for thin objects or when there are strong shadows for example.

[vict0rsch]

Hmm this isn't really good, is it? What's a representative inference @melisandeteng ? (I mean, not a worst nor best case scenario)

[vict0rsch]

Is this something we want in the test images?