Building the Context aggregation network model for testing on Indian Driving Dataset for Semantic Segmentation
- One of the main issues in semantic segmantation is to aggregate multiscale contextual information effectively. For this, a novel paradigm called Chained Context Agregation Module (CAM) has been proposed.CAM gains features of various spatial scales through chain-connected ladder-style information flows. CAM captures features of various scales at different levels and aggregates them in stages by the chain-connected ladder-style information flows.
- Semantic segmentation is a vital task in computer vision used to assign currosponding semantic labels to individual pixels in images and has applications in automatic driving, medical imaging etc.
- FCN's gain increasing receptive feild and high level contexts - through cascaded convolution and pooling layers. However, continuous downsampling process causes loss of spatial details, resulting in poor object delination.
- I developed the model architecture as explained in the paper and ran it for 70 but stopped at 53 epochs since the metrics stopped improving. Link: https://arxiv.org/pdf/2002.12041.pdf
- I got a final accuracy of Accuracy is 69.96%; Mean IoU as 46.84 and Dice Coefficient as 62.64 on my test dataset.