Nanodet: A 97FPS mobile phone target detection model without anchor point

YOLO's alternative, the 97FPS anchor-free target detection model NanoDet on mobile phones is now open source ~

1, what is Nanodet?

Nanodet is an ultra-fast lightweight anchor-free target detection model for mobile terminals, which is improved on the basis of FCOS model.

2. Performance comparison between 2.Nanodet and other models.

Running benchmark with NCNN on Huawei P30 requires only 10.23ms for a picture, which is three times faster than YoloV4-Tiny and six times smaller in parameters. Cocoa map (05:0.95) can reach 20.6. The model weight is only1.8m.

It can be seen that Nanodet can have the same texture as YoloV4-Tiny, with lower parameters and smaller model weight file.

3. the network structure diagram of 3.Nanodet.

The trunk is Shuffnet V2 1.0x, the convolution layer of the last layer is removed, the features with sampling rate of 8/ 16/32 times are extracted, and the features are input to PAN for multi-scale feature fusion.

4. the improvement of 4.Nanodet compared with FCOS.

① Probe head

FCOS:* * * * * * Enjoy the weight of the probe.

Nanodet: You don't * * enjoy the probe head, because the CPU is used for calculation on the mobile side, and * * enjoying the weight of the probe head will not accelerate the reasoning process; * * * Sharing the probe will further reduce its detection capability.

② Standardization mode

FCOS: Using Group Normalization

Nanodet: Batch normalization is used instead of group normalization, because the latter has many advantages over the former, but it has a disadvantage: batch normalization can integrate its parameters into convolution kernel during reasoning, and it can save time by omitting this step of calculation.

③ the size of the probe

FCOS: The size of the detector head is that four 256-channel convolution kernels are one branch, so boundary regression and classification are two branches, so a total of eight 256-channel convolution kernels are needed.

Nanodet: Reduce 4 convolution kernel groups to 2 convolution kernel groups. The number of channels was reduced from 256 to 96. Boundary regression and classification are shared by detectors, and then cut into two parts to predict boundary regression and classification respectively.

④ FPN layer

FCOS: There is no such module.

Nanodet: Based on FPN, improved to PAN. The original FPN uses 3*3 convolution after linear interpolation on the high-level feature map. However, Nanodet deleted the 3*3 convolution used after FPN linear interpolation, and only kept the 1* 1 convolution extracted from the backbone network elements to align the dimensions of the feature channels. At the same time, only top-down feature fusion is used at the multi-layer feature fusion end, while Pan not only uses top-down feature fusion, but also uses bottom-up feature fusion, and the method adopted is also linear interpolation.

5. Algorithm steps of 5.Nanodet

6. loss function of 6.Nanodet

7. advantages of 7.Nanodet

(1) is fast.

② The model parameter weight file is small.

8. disadvantages of 8.Nanodet

① The map is not high