Research Ideas
-
Competitive methods for WSSS from image-level labels: PRM and successor, PSA and successor
-
HSN-v2 ** CNN+Grad-CAM+CRF: end-to-end training with learnable BG/OT activations (ADP tuning), loss function to maximize single pixel confidence, global pixel confidence, minimize Grad-CAM overlapping, intra-segment visual heterogeneity (reduce CRF changes) (ADP training) ** CNN+Attention Mechanism+Grad-CAM+CRF: end-to-end training, with progressive weighting with attention mechanism ** CNN+Multi-Scale Grad-CAM+CRF: end-to-end training, with amalgamation of Grad-CAM at multiple scales and/or with guided backpropagation to get finer details ** CNN+Grad-CAM+Random Walk+CRF: end-to-end training, with random walk outward/inward from seeds
-
Multi-Label Pixel-Preserving Data Augmentation: implement training that uses Grad-CAM progressively to eliminate GT labels that are cut off by image augmentation transformations
-
Reading ** Boykov and Partial Cross-Entropy Loss
-
Evaluation ** Other WSSS methods: lots of Grad-CAMs overlap in pathology, but not in natural scene images