This project is the Naver Boost Camp CV11 team's submission code for the trash object detection competition. Given an image containing a garbage object, it is a matter of specifying the location of the garbage and classifying the class.
- OS : Linux Ubuntu 18.04.5
- GPU : Tesla V100 (32GB)
├─eda
├─ensemble
├─mmdetection
├─mmdetection3
├─UniverseNet
├─yolov8
├─multilabel_kfold.py
└─streamlit
pip install -r requirements.txt
You can enter the multilabel_kfold.py folder path and enter the command or enter the multilabel_kfold directly into the command
python3 level2_objectdetection-cv-11/multilabel_Kfold.py --K {kfold split count}
-
Move the path to the tools folder where the train.sh file is located
-
Write python3 main.py command in train.sh file
python3 main.py --model {.py name} --folder {folder_name} --resize {size} --max_epoch {epoch} --inference_epoch {best/latest}
-
Run
nohup sh train.sh
Metric : mAP score
augmentation | mAP50 | General trash(0) | Plastic(5) | total |
---|---|---|---|---|
Normalize | 0.346174 | 0.173843 | 0.280453 | 0.80047 |
HorizontalFlip | 0.317365 | 0.17596 | 0.254375 | 0.7477 |
Mosaic | 0.328192 | 0.16313 | 0.252239 | 0.743561 |
RGBShift | 0.324352 | 0.152785 | 0.248293 | 0.72543 |
MedianBlur | 0.317173 | 0.145847 | 0.25099 | 0.71401 |
HueSaturation | 0.310931 | 0.163762 | 0.238093 | 0.712786 |
Cutout | 0.311202 | 0.157962 | 0.2429 | 0.712064 |
CLAHE | 0.308945 | 0.158813 | 0.240687 | 0.708445 |
JpegCompression | 0.32155 | 0.16028 | 0.22024 | 0.70207 |
RandomBrightnessContrast | 0.312545 | 0.144382 | 0.238167 | 0.695094 |
Multiresize | 0.29278 | 0.15752 | 0.235295 | 0.685595 |
Metric : mAP score
Library | TYPE | Method | Backbone | Neck | Datasets | Scheduler | Runtime | Optimizer | mAP(public) |
---|---|---|---|---|---|---|---|---|---|
mmdetection2 | 2stage | Cascade RCNN | Swin transformer base | FPN | albu_coco_detection | schedule | default_runtime | AdamW | 0.6671 |
UniverseNet | 1stage | UniverseNet | Res2Net_101 | FPN / SEPC | albu_coco_detection | schedule_20e | default_runtime | AdamW | 0.61 |
UniverseNet | 1stage | ATSS | Swin transformer large | FPN / Dyhead | albu_coco_detection | schedule_20e | default_runtime | AdamW | 0.6237 |
UniverseNet | 2stage | GFLv2 | PVT_v2 | FPN | albu_coco_detection | schedule_2x | default_runtime | SGD | 0.5693 |
UniverseNet | 1stage | TOOD | Swin transformer v2 tiny | FPN | albu_coco_detection | schedule_adamw | default_runtime | AdamW | 0.54 |
YOLOv8 | 1stage | YOLOv8_l | Darknet | - | recycle.yaml | - | - | - | 0.3822 |
- Final submission : Public : 0.6878(5th) / Private : 0.6720(5th)