- Object Detection

Using automatic compression allows for easy model compression without a deep understanding of compression. Advanced compression offers higher flexibility in settings and may yield better compression results.

'Compressed' Type: Automatic Compression
'Compressed(Adv.)' Type: Advanced Compression

All of the original and compressed models can be downloaded easily on the Model Compressor Model Zoo.

PyTorch

Model	Best Practice	Type	Dataset	mAP(0.5) (%)	mAP(0.5:0.95)(%)	FLOPs (M)	Params (M)	Latency (ms)	Model Size (MB)
YOLOX		Original	COCO	68.0	49.7	156006.20	54.21	12239.46	207.37
YOLOX	Google Colab	Compressed-1	COCO	67.16 (-0.84)	48.64 (-1.06)	101804.06 (1.53x)	19.96 (2.7x)	8502.72 (1.44x)	76.61 (2.7x)
YOLOX	Google Colab	Compressed-2	COCO	61.43 (-6.57)	43.23 (-5.47)	38607.03 (4.04x)	4.93 (11.0x)	4235.37 (2.89x)	19.17 (10.80x)

The model’s latency is measured on Raspberry Pi 4B (1.5GHz ARM Cortex).
Options: FP32, ONNX runtime

TensorFlow-Keras

Model	Best Practice	Type	Dataset	mAP(0.5) (%)	mAP(0.5:0.95)(%)	FLOPs (M)	Params (M)	Latency (ms)	Model Size (MB)
YOLOv4		Original	PASCAL VOC	82.22	-	61871.82	65.32	64318.70	262.90
YOLOv4	Google Colab	Compressed-1	PASCAL VOC	87.23 (+5.01)	-	11459.69 (5.4x)	10.59 (6.17x)	28651.70 (2.16x)	44.12 (5.96x)
YOLOv4	Google Colab	Compressed-2	PASCAL VOC	87.91 (+5.69)	-	14442.96 (4.28x)	10.71 (6.1x)	28976.40 (2.14x)	44.36 (5.93x)

YOLOv4 model with EfficientNet B1 based backbone.
The model’s latency is measured on Raspberry Pi 4B (1.5GHz ARM Cortex).
Options: FP32, TFLite