Computer vision

Improving object detection models

Dropblock on top of each convolution layer increased the overall performance on imagenet by 1.6%. Note that we should keep the probability high in the inital (close to 1) and reduce it to 0.85 over the period of training. check if dropblock is der in your network or not.
Use CIOU>DIOU>GIOU>IOU>(L1&L2 loss) for bbox regression. In anchor based models, we saw relative increase in mAP by 10% (3-5% mAP).

[30-12-2022] ATSS - Adaptive Training Sample Selection (ATSS) 👷
[29-12-2022] FCOS - Object detection without Anchor boxes. 👷
[19-09-2022] Vision Transformer - Understanding vision transfomer in 6 simple steps.
[01-09-2022] Cutout - A new generalization method, performs better than dropout.
[22-08-2022] CSPNet - Make resnet more efficient.
[15-08-2022] structural reparameterization - Get VGG style speed with Resnet kind of architectures.
[10-08-2022] CIOU>DIOU>GIOU>IOU>(L1&L2 loss) - bbox_regression_improvements.ipynb