DOWNLOAD THE MODEL FILE:
yolov9-s-relu6-best_320_int8_edgetpu.tflite
Model weights file for YOLO v9 "s" with resolution 320x320 compiled to run on Google Coral using the modified edgetpu detector plug-in code.
This version uses a different ordering for the boxes tensor which is better suited to the EdgeTPU limits and for post-processing, which improves the detection latency to 11ms (from 13ms).
17 classes of objects can be detected (a subset of COCO's 80 classes)
Model performance after 10 epochs of fine-tuning.
epoch, train/box_loss, train/cls_loss, train/dfl_loss, metrics/precision, metrics/recall, metrics/mAP_0.5,metrics/mAP_0.5:0.95, val/box_loss, val/cls_loss, val/dfl_loss, x/lr0, x/lr1, x/lr2
9, 1.47, 1.3719, 1.649, 0.75295, 0.65368, 0.72426, 0.54616, 0, 0, 0, 0.00208, 0.00208, 0.00208