From a653cc4c2c77e467fa4fb05fdd390f62593dbf7a Mon Sep 17 00:00:00 2001 From: zoooo0820 Date: Fri, 13 May 2022 03:52:07 +0000 Subject: [PATCH] more tip for training custimized dataset, test=document_fix --- configs/keypoint/README.md | 2 +- configs/keypoint/README_en.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/configs/keypoint/README.md b/configs/keypoint/README.md index 6db5b975d..771706d3b 100644 --- a/configs/keypoint/README.md +++ b/configs/keypoint/README.md @@ -208,7 +208,7 @@ upper_body_ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] #上半身对应关键点i #### 3、数据准备注意 - 训练数据请按coco数据格式处理。需要包括关键点[Nx3]、检测框[N]标注。 -- 请注意area>0，area=0时数据会被过滤掉。 +- 请注意area>0，area=0时数据在训练时会被过滤掉。此外，由于COCO的评估机制，area较小的数据在评估时也会被过滤掉，我们建议在自定义数据时取`area = bbox_w * bbox_h`。 如有遗漏，欢迎反馈 diff --git a/configs/keypoint/README_en.md b/configs/keypoint/README_en.md index 486d0d4ab..75881a979 100644 --- a/configs/keypoint/README_en.md +++ b/configs/keypoint/README_en.md @@ -214,7 +214,7 @@ For more configs, please refer to [KeyPointConfigGuide](../../docs/tutorials/Key #### 3、Note for data preparation - The data should has the same format as Coco data, and the keypoints(Nx3) and bbox(N) should be annotated. -- please set "area">0 in annotations files otherwise it will be skiped while training. +- please set "area">0 in annotations files otherwise it will be skiped while training. Moreover, due to the evaluation mechanism of COCO, the data with small area may also be filtered out during evaluation. We recommend to set `area = bbox_w * bbox_h` when customizing your dataset. ## BenchMark -- GitLab