Computer Vision News - February 2018

RSIP Vision ’s CEO Ron Soferman has launched a series of lectures to provide a robust yet simple overview of how to ensure that computer vision projects respect goals, budget and deadlines. This month we learn about a special situation with Deep Learning: Thriving for More than 90% . It’s another tip for Project Management in Computer Vision . As an R&D manager or team leader , you can very soon find yourself in a situation where your team shows good results from the POC stage of a Deep Learning project . The results might be very encouraging, in the range of 85%-90% success rate ( Precision / Recall , i.e. Positive Predictive Value / Sensitivity). Still, you might wonder about the ways to reach the next level. Here we propose several directions to consider: 1. Different network architectures . 2. More data for training : when this is applicable and it does not cause excessive delay to the work schedule. 3. Add pre-processing : sometimes it can emphasize the important features; rotate and resize to canonical form to enable smaller training set. 4. Add post-processing : removing false results. 5. Sort out errors from the training sets : false negatives and false positives by human annotation, whether you used a Mechanical Turk or any other method. 6. Cont. : in some [lucky] cases, network performance may already be better than the human labels, which will need to be corrected! 7. Different resolution might play a crucial role: we prefer low resolution for higher speed; but we might need the details for the classification task. 8. Combining 2 or more schemes , that can contribute - when taken together - to produce better results. 9. Examine different training objects : One can train on part of the object that is more prominent and less ambiguous than the whole object. 10. Fine-tune hyper-parameters : try different learning rates, mini-batch size and momentum. 11. Combining different loss functions : this is generally helpful. 12. Verification of data augmentation results : it must look authentic and represent the multitude of object appearance. If it fails, we train the net for our artifacts. 13. Visually examine a sample of specific failed and successful cases. You may find specific issues or situations that call for a solution which is easily understandable by a human observer. 14. Use better input data : for instance, we, at RSIP Vision , have seen that coronal planes work better than sagittal planes for disk segmentation in spinal scans. Deep Learning: Thriving for More than 90% Management Computer Vision News Project Management Tip 27