YonghaoHe/LFD-A-Light-and-Fast-Detector

Update History

• 2021.06.25 Some new experiments are added for your reference.
• 2021.03.30 We maintain an Experiments wiki page to show ablation studies for your reference. Maybe these experiments are valuable for you to make proper decisions.
• 2021.03.16 INT8 inference is updated. Check timing_inference_latency.py and predict_tensorrt.py for reference.
• 2021.03.09 LFD now is formally released!!! Any questions and problems are welcome.

1. Introduction

In this repo, we release a new One-Stage Anchor-Free Detector named LFD. LFD completely surpasses the previous LFFD in most aspects. We are trying to make object detection easier, explainable and more applicable. With LFD, you are able to train and deploy a desired model without all the bells and whistles. Eventually, we hope LFD can be as popular as YOLO series for the industrial community in the future.

1.1 New Features

Compared to LFFD, LFD has the following features:

• implemented using PyTorch, which is friendly for most guys (LFFD is implemented using MXNet)
• support multi-class detection rather than single-class detection (LFFD is only for single-class)
• higher precision and lower inference latency
• we maintain a wiki (highly recommended) for you to fully understand LFD and master the code
• the performance of LFD has been proved in many real-world applications
• create your desired models with satisfactory model size and inference latency, train from scratch on your own datasets,

1.2 Performance Highlights

Before dive into the code, we present some performance results on two datasets, including precision and inference latency.

Dataset 1: WIDERFACE (single-class)

Accuracy on val under the SIO evaluation schema proposed in LFFD

• v2 is from LFFD, you can check it in LFFD repo.
• for fairy comparison, WIDERFACE-L/M/S/XS have the similar detection range with v2, namely [4, 320] vs [10, 320], but WIDERFACE-L/M/S/XS have different network structures.
• great improvement on Hard Set.

Inference latency

Platform: RTX 2080Ti, CUDA 10.2, CUDNN 8.0.4, TensorRT 7.2.2.3

• batchsize=1, weight precision mode=FP32

the results of v2 is directly get from LFFD, the Platform condition is slightly different from here.

• batchsize=1, weight precision mode=FP16

It can be observed that FP16 mode is evidently faster than FP32 mode. So in deployment, FP16 is highly recommended if possible.

• batchsize=1, weight precision mode=INT8

CAUTION: - means results are not available due to out of memory while calibrating

Dataset 2: TT100K (multi-class----45 classes)

Precision&Recall on test set of TT100K[1]

We use only train split (6105 images) for model training, and test our models on test split (3071 images). In [1], authors extended the training set: Classes with between 100 and 1000 instances in the training set were augmented to give them 1000 instances. But the augmented data is not released. That means we use much less data than [1] used for training. However, as you can see, our models can still achieve better performance. Precision&Recall results of [1] can be found in it's released code folder: code/results/report_xxxx.txt.

Inference latency

Platform: RTX 2080Ti, CUDA 10.2, CUDNN 8.0.4, TensorRT 7.2.2.3

• batchsize=1, weight precision mode=FP32

• batchsize=1, weight precision mode=FP16

• batchsize=1, weight precision mode=INT8

CAUTION: - means results are not available due to out of memory while calibrating

2. Get Started

2.1 Install

Prerequirements

• python => 3.6
• albumentations => 0.4.6
• torch => 1.5
• torchvision => 0.6.0
• cv2 => 4.0
• numpy => 1.16
• pycocotools => 2.0.1
• pycuda = 2020.1
• tensorrt = 7.2.2.3 (corresponding cudnn = 8.0)

All above versions are tested, newer versions may work as well but not fully tested.

Build Internal Libs

In the repo root, run the code below:

python setup.py build_ext

Once successful, you will see: ----> build and copy successfully!

if you want to know what libs are built and where they are copied, you can read the file setup.py.

Build External Libs

• Build libjpeg-turbo

  1. download the source code v2.0.5

  2. decompress and compile:

     cd [source code]
mkdir build
cd build
cmake ..
make

     make sure that cmake configuration properly

  3. copy build/libturbojpeg.so.x.x.x to lfd/data_pipeline/dataset/utils/libs

Add PYTHONPATH

The last step is to add the repo root to PYTHONPATH. You have two choices:

1. permanent way: append export PYTHONPATH=[repo root]:$PYTHONPATH to the file ~/.bashrc
2. temporal way: whenever you want to code with the repo, add the following code ahead:
   1. import sys
   2. sys.path.append('path to the repo')

Until now, the repo is ready for use. By the way, we do not install the repo to the default python libs location (like /python3.x/site-packages/) for easily modification and development.

Docker Installation

please check here for more details, thanks to @ashuezy.

2.2 Play with the code

We present the details of how to use the code in two specific tasks.

• face detection on WIDERFACE README.md
• traffic sign detection on TT100K README.md

Besides, we describe the structure of code in wiki.

Acknowledgement

• very much thankful for PyTorch framework.
• we learn a lot and reuse some basic code from mmdetection, thanks for the great work.
• thanks for some third-party libs like albumentations, turbojpeg.

Citation

If you find the repo is useful, please cite the repo website directly.