Owner: Peizhao Zhang (stzpz@fb.com)

• MaskRCNN2GO (bbox + segmentation) with 81 classes
• float32 and int8 models
• Trained on COCO 2014 dataset
• int8 model fine-tuned with fake-quantization

• Evaluation dataset: COCO 2014 minival
• Metric: mAP[IoU=0.50:0.95] and latency (in microseconds)
• Proposals: 3000/100 (pre/post nms)
• The backend for evaluation is QNNPACK
• The unit of mensured latency is microseconds
• Latency is evaluated on Samsung S8 (SM-G950U-7.0-24)

• data (1, 3, H, W), min(H, W) = 320, BGR in range [0, 255]
• im_info (1, 3) [scaled_height, scaled_width, scale]

• Download COCO 2014 minival dataset
• OpenCV must be installed for image preprocessing
• Run run.sh [coco_dir] [model]
• The model argument can be either fp32 for float 32 or int8

• f93520960
• f96110081:934610037

Thanks a lot for the help from Carole-Jean Wu, Fei Sun, Yiming Wu, Yanghan Wang and Zhizhen Qin.