Cell segmentation and multi-object tracking using deep learning
The project aims at solving object detection and tracking of uni-cellular amoebas from videos. The
objective is to employ a deep learning architecture which can simultaneously detect the cells and
associate the cells over time. The methods are generally employed in two steps: the first step detects
cellular objects in every frame of a video (2D detection). The second step establishes association
between the detected objects in consecutive frames to obtain their tracks.
Various deep learning architectures are available for detection and segmentation. While the
association is performed generally via optimization methods using hand-crafted features. Some recent
works [1,2,3] has designed deep-learning models itself to simultaneously solve the segmentation and
object association problem. In general, these methods employ a sub-network for object detection and
use the neural-network features in a second sub-network to generate a correspondence map between
objects in corresponding frames. This correspondence map between detected objects is used to
associate objects in the current frame with that of objects in previous frames for reliable tracking over
1. Ground truth data generation for the association problem. The object detection ground truth will be
provided. The ground truth determining the links between objects over time needs to be generated.
2. The second part of the project involves using one of the architectures form the reference to perform
training on the ground truth data and use the trained model on the test dataset.
1. Sun, ShiJie, et al. "Deep affinity network for multiple object tracking." IEEE transactions on pattern
analysis and machine intelligence (2019).
2. Schulter, Samuel, et al. "Deep network flow for multi-object tracking." Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition. 2017.
3. Zhang, Wenwei, et al. "Robust multi-modality multi-object tracking." Proceedings of the IEEE
International Conference on Computer Vision. 2019.