v1.1.0

ObjectDetector v1.1.0

Diff since v1.0.0

Fixes

• detect_thresh was erroneously filtering the objectness score (5th row) (aka the bbox confidence), not the class confidence (end-2 row) (the product of class prediction and objectness score), as it now uses.

Improvements

• YOLO v3-spp, v4, v4-tiny, v7, v7-tiny now supported
• All models v3+ now have result parity with AlexeyAB/darknet
• All models v3+ tested directly against Darknet.jl
• draw_boxes now draws class-colored boxes and confidence labels

Merged pull requests:

• Add yolov4 and yolov7 model artifacts (#119) (@IanButterworth)
• Get YOLOv4 and v7 working and tested against darknet (#120) (@IanButterworth)
• Testing improvements. Remove precompile workload. (#121) (@IanButterworth)
• Test pretrained first (#122) (@IanButterworth)
• Test all working models against Darknet. Fix padding tests. (#123) (@IanButterworth)
• Sort the refs (#125) (@IanButterworth)
• Fix conf thresholding. Compare on GPU. Minor code improvements (#126) (@IanButterworth)
• draw_boxes: make boxes class colored. add conf labels (#127) (@IanButterworth)

Closed issues:

• Yolo V4-tiny (#49)
• Choice of bounding box color (#80)
• Test results differ compared to darknet for nonsquare images (#124)