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

Name		Name	Last commit message	Last commit date
Latest commit History 32 Commits
archive_course		archive_course
images		images
.gitignore		.gitignore
ATSS.ipynb		ATSS.ipynb
CSPNet.ipynb		CSPNet.ipynb
Clip.ipynb		Clip.ipynb
FCOS.ipynb		FCOS.ipynb
Finding_optimal_anchors.ipynb		Finding_optimal_anchors.ipynb
README.md		README.md
Understanding_object_detection_metrics_and_errors.ipynb		Understanding_object_detection_metrics_and_errors.ipynb
bbox_regression_improvements.ipynb		bbox_regression_improvements.ipynb
cutout.ipynb		cutout.ipynb
settings.ini		settings.ini
setup.py		setup.py
structural_reparameterization.ipynb		structural_reparameterization.ipynb
vision_transformers.ipynb		vision_transformers.ipynb