tahirishaq10/widerface_kitti

Install TAO Toolkit using pip and make sure to pull its docker container with GPU runtime (if using Colab or similar service) otherwise, the operation will fail. It is a relatively large image (21GB) as a result it will a while to download.

Training YoloV4 Tiny with widerface dataset using Nvidia TAO toolkit. TAO Yolov4 Tiny requires the input image shape to be a multiple of 32 therefore, the images were resized to 768 x 768 and were also converted to PNG format.

Could not find the pretrained Yolov4 tiny model using NCG CLI therefore, had to download manually from here.

There were only 6 anchor shapes/tuples were generated instead of the mentioned 9. Moreover, in tao_tutorials/notebooks/tao_launcher_starter_kit/yolo_v4_tiny/specs/yolo_v4_tiny_train_kitti.txt along with other kitti files also only had mid and big ancher shapes. The anchors generated for custom dataset

Extra class mapping also had be commented out/removed from kitti files. For all other parameters, default values were used.

References

• WIDERFACE
• Nvidia TAO tutorial
• KITTI
• TAO Objected Detection Pretrained model
• Yolov4 DeepStream
• DeepStream YOLO