scrfd 问题

[yangjian1218]

trafficstars

问题1:scrfd模型 bnkps 跟gnkps是什么意思? 问题2: scrfd怎么多batch输入, 我把onnx改为动态batch,可是输入多张图片进去,输出维度跟一张图片一样.这是为什么

[yangjian1218]

错了 输出维度不一样, 后处理就只要按batch平均切分么

[SthPhoenix]

问题1:scrfd模型 bnkps 跟gnkps是什么意思? 问题2: scrfd怎么多batch输入, 我把onnx改为动态batch,可是输入多张图片进去,输出维度跟一张图片一样.这是为什么

1. KPS - means those models extracts facial landmarks in addition to bounding boxes. BN - means model use BatchNorm layers and GN - GroupNorm (InstanceNorm). Default DeepInsight models have bug with detecting large faces, changing BN to GN during training helped solve this, though at cost of some accuracy and memory consumption.

2. Hard to tell not knowing your actual code, in this repo SCRFD batch processing is supported only for TensorRT backend.

错了 输出维度不一样, 后处理就只要按batch平均切分么

I'm not sure I understand you, Google Translate gives a bit weird translation. You can refer to scrfd.py for post processing example.