jungyeon working directory

BEFORE running the code, you have to download Kitti dataset! And then, change the dataset directory in main.py. Download

Setting Virtual Environment with Anaconda

conda create -n [VIR_ENV_NAME] python==3.8

conda activate [VIR_ENV_NAME]

pip install -r requirements.txt

python main.py

If you can run main.py, the setting is done.

Active Learning for EFFICIENT Learning

Goals: random sampling VS. confidence VS. LL4AL

0. SSD(ResNet50) + Kitti(Object detection) + Confidence

1. SSD(ResNet50) + Kitti(Object detection) + Learning loss

   • Learning loss code Github
   • SSD(ResNet50) Github
   • Wandb Github
   • Kitti parameter Github
   • maybe later : PyTorch 1.5.0 이상 버전에서 model weight update와 관련된 inplace operation이 수정되면서 zero_grad -> backward -> step의 과정이 최적화하고자 하는 변수의 순서에 맞춰서 수행되어야 해결되는 것으로 알고 있습니다.

   

---

Reference

SSD

1. [논문] SSD: Single Shot Multibox Detector 분석
2. 갈아먹는 Object Detection [6] SSD
3. mAP 계산하기
4. Object Detection
5. Object-Detection-Object-Detection-튜토리얼
6. Object detection: speed and accuracy comparison (Faster R-CNN, R-FCN, SSD, FPN, RetinaNet and YOLOv3)
7. Deep Learning for Generic Object Detection: A Survey
8. a-PyTorch-Tutorial-to-Object-Detection
9. nvidia_deeplearningexamples_ssd

AL

1. Learning Loss for Active Learning : base
2. Learning to Rank for Active Learning: A Listwise Approach

KITTI

1. kitti vis code : random_vis.py 코드
2. kitti label : 라벨링 정보
3. kitti dataloader pytorch : dataset.py

---

ETC

• argparse

if time permits

+RL

• Learning how to Active Learn: A Deep Reinforcement Learning Approach

+SSL

• Multiple Instance Active Object Detection (MI-AOD)
• Unbiased Teacher for Semi-Supervised Object Detection