This repo is customed for VisDrone.

Last update: Jul 17, 2022

Overview

Object Detection for VisDrone(无人机航拍图像目标检测)

My environment

1、Windows10 (Linux available)
2、tensorflow >= 1.12.0
3、python3.6 (anaconda)
4、cv2
5、ensemble-boxes(pip install ensemble-boxes)

Datasets(XML format for training set)

(1).Datasets is available on https://github.com/VisDrone/VisDrone-Dataset
(2).Please download xml annotations on Baidu Yun (提取码: ia3f), or Google Drive, and configure it in ./core/config/cfgs.py
(3).You can also use ./data/visdrone2xml.py to generate your visdrone xml files, modify the path information.

training-set format:

├── VisDrone2019-DET-train
│     ├── Annotation(xml format)
│     ├── JPEGImages

Pretrained Models(ResNet50vd, 101vd)

Please download pretrained models on Baidu Yun (提取码: krce), or Google Drive, then put it into ./data/pretrained_weights

Train

Modify the parameters in ./core/config/cfgs.py
python train_step.py

Eval

Modify the parameters in ./core/config/cfgs.py
python eval_visdrone.py, it will get txt format file, then use official matlab tools to eval the final results.
python eval_model_ensemble.py. Before the running of this file, you should set NORMALIZED_RESULTS_FOR_MODEL_ENSEMBLE=True in cfgs.py and then run eval_visdrone.py to get normalized txt result.

Visualization

Modify the parameters in ./core/config/cfgs.py
python image_demo.py, it will get visualized results.

Visualized Result (multi-scale training+multi-scale testing)

Test Result(Validation set)：

1. ResNet50-vd

Name	maxDets	Result(s/m)
Average Precision (AP) @( IoU=0.50:0.95)	maxDets=500	31.26%/35.1%
Average Precision (AP) @( IoU=0.50 )	maxDets=500	56.44%/60.29%
Average Precision (AP) @( IoU=0.75 )	maxDets=500	30.13%/35.42%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 1	0.78%/0.58%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 10	6.62%/6.05%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=100	38.21%/40.99%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=500	48.41%/53%

"s" means single-scale training + single-scale testing; "m"means multi-scale training + multi-scale testing

2. ResNet101-vd

Name	maxDets	Result(s/m)
Average Precision (AP) @( IoU=0.50:0.95)	maxDets=500	31.7%/35.98%
Average Precision (AP) @( IoU=0.50 )	maxDets=500	56.94%/61.64%
Average Precision (AP) @( IoU=0.75 )	maxDets=500	30.59%/36.13%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 1	0.67%/0.61%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 10	6.29%/6.13%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=100	38.66%/42.33%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=500	49.29%/53.68%

3. Model Ensemble (ResNet101-vd+ResNet50-vd)

Name	maxDets	Result
Average Precision (AP) @( IoU=0.50:0.95)	maxDets=500	36.76%
Average Precision (AP) @( IoU=0.50 )	maxDets=500	62.33%
Average Precision (AP) @( IoU=0.75 )	maxDets=500	37.41%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 1	0.59%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets= 10	6.06%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=100	42.57%
Average Recall (AR) @( IoU=0.50:0.95)	maxDets=500	54.53%

You can download trained weights(ResNet50vd, 101vd) on Baidu Yun (提取码: 9u9m), or Google Drive, then put it into ./saved_weights

Reference

1、https://github.com/DetectionTeamUCAS/Faster-RCNN_Tensorflow
2、https://github.com/open-mmlab/mmdetection
3、https://github.com/ZFTurbo/Weighted-Boxes-Fusion
4、https://github.com/kobiso/CBAM-tensorflow-slim
5、https://github.com/SJTU-Thinklab-Det/DOTA-DOAI
6、https://github.com/Viredery/tf-eager-fasterrcnn
7、https://github.com/VisDrone/VisDrone2018-DET-toolkit
8、https://github.com/YunYang1994/tensorflow-yolov3
9、https://github.com/zhpmatrix/VisDrone2018

This repo is customed for VisDrone.

Related tags

Overview

Object Detection for VisDrone(无人机航拍图像目标检测)

My environment

Datasets(XML format for training set)

Pretrained Models(ResNet50vd, 101vd)

Train

Eval

Visualization

Test Result(Validation set)：

1. ResNet50-vd

"s" means single-scale training + single-scale testing; "m"means multi-scale training + multi-scale testing

2. ResNet101-vd

3. Model Ensemble (ResNet101-vd+ResNet50-vd)

You can download trained weights(ResNet50vd, 101vd) on Baidu Yun (提取码: 9u9m), or Google Drive, then put it into ./saved_weights

Reference

Owner

SalFBNet: Learning Pseudo-Saliency Distribution via Feedback Convolutional Networks

ICLR 2021: Pre-Training for Context Representation in Conversational Semantic Parsing

Code for Universal Semi-Supervised Semantic Segmentation models paper accepted in ICCV 2019

The official code of "SCROLLS: Standardized CompaRison Over Long Language Sequences".

The open source code of SA-UNet: Spatial Attention U-Net for Retinal Vessel Segmentation.

Implement some metaheuristics and cost functions

[ICCV2021] 3DVG-Transformer: Relation Modeling for Visual Grounding on Point Clouds

AFLNet: A Greybox Fuzzer for Network Protocols

Implementation of the 😇 Attention layer from the paper, Scaling Local Self-Attention For Parameter Efficient Visual Backbones

PyTorch implementation of DeepUME: Learning the Universal Manifold Embedding for Robust Point Cloud Registration (BMVC 2021)

Resources related to EMNLP 2021 paper "FAME: Feature-Based Adversarial Meta-Embeddings for Robust Input Representations"

The implementation for paper Joint t-SNE for Comparable Projections of Multiple High-Dimensional Datasets.

InvTorch: memory-efficient models with invertible functions

Implementation of the CVPR 2021 paper "Online Multiple Object Tracking with Cross-Task Synergy"

Code for the paper One Thing One Click: A Self-Training Approach for Weakly Supervised 3D Semantic Segmentation, CVPR 2021.

This repository contains the code and models necessary to replicate the results of paper: How to Robustify Black-Box ML Models? A Zeroth-Order Optimization Perspective

Continuous Augmented Positional Embeddings (CAPE) implementation for PyTorch

Pipeline code for Sequential-GAM(Genome Architecture Mapping).

Supervised Contrastive Learning for Downstream Optimized Sequence Representations

Linear image-to-image translation