Paper implementation of Cross Image Pixel Contrast for Semantic Segmentation (ICCV 21')

Code can be integrated in your semantic segmentation training pipeline using the following steps

1. Initialise ContrastiveLearner class for training and validation respectively

   contrastive_learner_train = ContrastiveLearner(region_memory_bank_size = training_set_size,
                                                    pixel_memory_bank_size = 10 * training_set_size,
                                                    num_classes = num_classes)
   
   contrastive_learner_val = ContrastiveLearner(region_memory_bank_size = validation_set_size,
                                                pixel_memory_bank_size = 10 * validation_set_size,
                                                num_classes = num_classes)
   

2. Build memory bank inside training loop (after forward pass on a batch is done)

   contrastive_learner_train.build_memory_bank(features = features.detach(),index = index, label = y.detach())

3. Calculate contrastive loss inside training loop

   contrastive_loss = contrastive_learner_train.compute_loss(preds,features,y)