In this talk, we’ll explore an experimental process which maximizes the performance of object detection models, with almost no retraining effort. Since fine-tuning and retraining deep models is expensive, this talk focuses on extracting the most accuracy out of small datasets. The core idea of the talk is that mistakes in object detection can often be tracked down to sub-optimal stratification of training and testing datasets. By leveraging different train / test stratification strategies to expose these mistakes correctly, we can show statistically and numerically significant improvements in our metrics. Stratification for classification and regression problems is easy, but it becomes very difficult for compound problems like object detection. This talk, in part, addresses this issue.
Every mistake made by an object detection model can be characterised with elementary statistical distributions. This talk presents a set of experimental methods that together leverage these statistics to squeeze the last bit of accuracy out of an object detection model. These statistics are typically derived from the training data itself, and therefore these methods are very cheap – they don’t need any complex or expensive operations like image augmentation, model retraining or regularization.
During the talk, we will also take a closer look at various benchmark improvements that this method has been able to achieve over some well known object detection problems. Once this process has been adopted, conventional model improvement techniques can only improve model performance even further.
Key Takeaways:
- Maximize object detection performance with minimal retraining effort.
- Address sub-optimal stratification for improved object detection accuracy.
- Leverage elementary statistical distributions to optimize model performance.
- Achieve benchmark improvements in well-known object detection problems.
- Cost-effective approach without complex operations or retraining.
