This paper presents two novel approaches to learning swarming collective motion from vision data using both Reinforcement Learning and Imitation Learning. Both methods are trained with an identical dataset, network architecture and visual observation input to compare their ability to generate collective behaviours. A simulated environment is developed in Gazebo with a group of three ground vehicles performing flocking motion to validate the experiments. Comparative results reveal the characteristics of each method, including group, order and violation metrics. This study provides insights into the trade-offs between stability and adaptability in learning-based swarm control, offering guidance for future vision-based swarming systems.