第一次打比赛还是很不容易，感谢队友[@milleniums](https://github.com/milleniums). 一起合作
## 主要提分点
1、Mixup
2、DCN与多尺度训练测试
3、global context ROI
4、旋转数据增强
## A榜消融实验结果
<table>
<tr>
<th>Backbone</th>
<th>DCN</th>
<th>MS</th>
<th>Mixup</th>
<th>RandomRotate90</th>
<th>GC</th>
<th>mAP</th>
</tr>
<tr>
<th>ResNet50-FPN</th>
<th></th>
<th></th>
<th></th>
<th></th>
<th></th>
<th>baseline</th>
</tr>
<tr>
<th>ResNeXt101-FPN</th>
<th>✓</th>
<th>✓</th>
<th></th>
<th></th>
<th></th>
<th>baseline+3.35%</th>
</tr>
<tr>
<th>ResNeXt101-FPN</th>
<th>✓</th>
<th>✓</th>
<th>✓</th>
<th></th>
<th></th>
<th>baseline+4.25%</th>
</tr>
<tr>
<th>ResNeXt101-FPN</th>
<th>✓</th>
<th>✓</th>
<th></th>
<th></th>
<th>✓</th>
<th>baseline+4.36%</th>
</tr>
<tr>
<th>ResNeXt101-FPN</th>
<th>✓</th>
<th>✓</th>
<th></th>
<th>✓</th>
<th>✓</th>
<th>baseline+4.54%</th>
</tr>
<tr>
<th>ResNeXt101-FPN</th>
<th>✓</th>
<th>✓</th>
<th></th>
<th>✓</th>
<th>✓</th>
<th>baseline+4.69%</th>
</tr>
</table>
## 代码环境及依赖
参考requirement.txt
- **预训练模型下载**
- 下载mmdetection官方开源的htc的[resnext 64x4d 预训练模型](https://s3.ap-northeast-2.amazonaws.com/open-mmlab/mmdetection/models/htc/htc_dconv_c3-c5_mstrain_400_1400_x101_64x4d_fpn_20e_20190408-0e50669c.pth)
## 依赖安装及编译
- **依赖安装编译**
1. 安装 pytorch
conda install pytorch=1.1.0 torchvision=0.3.0 cudatoolkit=10.0 -c pytorch
2. 安装其他依赖
pip install cython && pip --no-cache-dir install -r requirements.txt
3. 编译cuda op等：
python setup.py develop
## 模型训练及预测
- **训练**
x101_64x4d (htc pretrained):
chmod +x tools/dist_train.sh
./tools/dist_train.sh configs/underwater/cas_x101/cascade_rcnn_x101_64x4d_fpn_1x.py 4
(上面的4是gpu数量，请自行修改,另外根据config对应文件目录进行修改) 训练过程文件及最终权重文件均保存在config文件中指定的workdir目录中
- **预测**
x101_64x4d (htc pretrained):
chmod +x tools/dist_test.sh
./tools/dist_test.sh configs/underwater/cas_x101/cascade_rcnn_x101_64x4d_fpn_1x.py workdirs/cas_x101_64x4d_fpn_htc_1x/latest.pth 4 --json_out results/cas_x101.json
1. 预测结果文件会保存在 /results 目录下
2. 转化mmd预测结果为提交csv格式文件：
python tools/post_process/json2submit.py --test_json cas_x101.bbox.json --submit_file cas_x101.csv
最终符合官方要求格式的提交文件cas_x101.csv 位于 submit目录下
## Reference
[Baseline @郑烨](https://github.com/zhengye1995/underwater-objection-detection).
和鲸社区Kesci 2020年全国水下机器人（湛江）大赛水下目标检测算法赛（光学）三等奖方案-单模.zip
版权申诉
139 浏览量
2023-11-05
23:59:07
上传
评论
收藏 25.45MB ZIP 举报
和鲸社区Kesci 2020年全国水下机器人（湛江）大赛水下目标检测算法赛（光学）三等奖方案-单模.zip （249个子文件） 
soft_nms_cpu.cpp 415KB deform_conv_cuda.cpp 29KB deform_pool_cuda.cpp 4KB roi_align_cuda.cpp 3KB roi_pool_cuda.cpp 3KB masked_conv2d_cuda.cpp 3KB nms_cpu.cpp 2KB sigmoid_focal_loss.cpp 2KB compiling_info.cpp 1KB nms_cuda.cpp 575B deform_conv_cuda_kernel.cu 42KB deform_pool_cuda_kernel.cu 16KB roi_align_kernel.cu 11KB roi_pool_kernel.cu 7KB sigmoid_focal_loss_cuda.cu 6KB nms_kernel.cu 5KB masked_conv2d_kernel.cu 5KB .DS_Store 12KB .DS_Store 8KB .DS_Store 6KB .DS_Store 6KB .DS_Store 6KB .DS_Store 6KB .DS_Store 6KB README.md 3KB transforms.py 35KB reppoints_head.py 27KB guided_anchor_head.py 25KB htc.py 22KB cascade_rcnn.py 21KB hrnet.py 19KB mean_ap.py 19KB resnet.py 18KB test_robustness.py 16KB fcos_head.py 16KB fovea_head.py 16KB grid_head.py 15KB generalized_attention.py 15KB flops_counter.py 14KB anchor_head.py 14KB two_stage.py 13KB senet.py 13KB deform_conv.py 12KB guided_anchor_target.py 12KB train.py 11KB bbox_head.py 11KB test_mixins.py 11KB test.py 10KB deform_pool.py 10KB grid_rcnn.py 9KB mask_scoring_rcnn.py 8KB max_iou_assigner.py 8KB cascade_rcnn_x101_64x4d_fpn_1x_Bbest.py 8KB cascade_rcnn_x101_64x4d_fpn_1x_Abest.py 8KB robustness_eval.py 8KB resnext.py 8KB coco_utils.py 8KB ssd_head.py 8KB transforms.py 8KB maskiou_head.py 7KB double_head_rcnn.py 7KB free_anchor_retina_head.py 7KB anchor_target.py 7KB fcn_mask_head.py 7KB inference.py 7KB setup.py 7KB base.py 7KB convfc_bbox_head.py 7KB coco_error_analysis.py 7KB iou_loss.py 7KB nas_fpn.py 6KB point_target.py 6KB ghm_loss.py 6KB analyze_logs.py 6KB decorators.py 6KB approx_max_iou_assigner.py 6KB formating.py 6KB recall.py 6KB iou_balanced_neg_sampler.py 6KB sampler.py 6KB conv_module.py 6KB point_assigner.py 6KB class_names.py 6KB eval_hooks.py 5KB ssd_vgg.py 5KB fpn.py 5KB double_bbox_head.py 5KB custom.py 5KB ga_rpn_head.py 5KB loading.py 5KB pascal_voc.py 5KB hooks.py 4KB coco.py 4KB single_level.py 4KB contextmanagers.py 4KB train.py 4KB rpn_head.py 4KB retina_sepbn_head.py 4KB detectron2pytorch.py 4KB context_block.py 4KB共 249 条
- 1
- 2
- 3
资源评论
