These systems are very sensitive to lighting changes. Potentially we could try to finetune with lighting augmentations to make them more robust since the training videos are shown in brightly lit areas. For now, we have been just adding more light to the demo.