OpenEarthMap Land Cover Mapping Few-Shot Learning Challenge
收藏资源简介：
***The challenge is over, please use this verion for post-challenge research. This version contains all the data and files, except the labels of the query-set of the testset that are withheld and researchers can submit the predictions at the challenge leaderboard for evaluation*** Overview This challenge is co-organized with the L3D-IVU 2024 CVPR workshop. The challenge is an extension of the OpenEarthMap benchmark dataset for a generalized few-shot semantic segmentation (GFSS) task. The challenge aims to evaluate and benchmark learning methods for few-shot semantic segmentation on the OpenEarthMap dataset to promote research on geoinformatics for social good. The motivation is to enable researchers to develop few-shot learning algorithms for high-resolution RS image semantic segmentation. Page https://cliffbb.github.io/OEM-Fewshot-Challenge/ Baseline The baseline model for the challenge is available here. Leaderboard The challenge leaderboard is opened for post-challenge research to enable researchers to evaluate the predictions on the query-set of the testset. Description The dataset has been designed for remote sensing few-shot learning, particularly, for GFSS tasks in land cover mapping. The dataset consists of only 408 samples from the original OpenEarthMap dataset for RS image semantic segmentation. It extends the original 8 semantic classes of the OpenEarthmap benchmark to 15 classes, which is split into 7:4:4 for train_base_class, val_novel_class, and test_novel_class disjointed sets, respectively (i.e., train_base_class ∩ val_novel_class ∩ test_novel_class = ∅). The 408 samples are also split into 258 as `trainset`, 50 as `valset`, and 100 as `testset`. The `trainset` is for pre-training a backbone network. It contains only the images and labels of the train_base_class split. Both the `valset` and the `testset` consist of a support set and a query set for a 5-shot with 4 novel classes and 7 base classes GFSS task. The `valset` and the `testset` contain the images and labels of the val_novel_class and the test_novel_class splits, respectively. The challenge is in two phases: development phase and evaluation phase. The `valset` is for the development phase and the `testsets` is for the evaluation phase. Both `valset` and `testset` have 20 image-label pair examples, 5-set examples for each of the 4 novel classes in the support set. The `valset` and the `testset` contain an additional 30 images and 80 images, respectively, in the query set, which are to be predicted using the 20 labelled images in their support set. The labels for each image in the support sets do not contain any of the train_base_class split. Also, in each 5-set examples, the labels contain only one novel class (i.e., one novel class per 5-set examples). However, in the `valset`, the labels for the images in the query set contain both train_base_class and val_novel_class; and in the `testset, the labels for the images in the query set contain both train_base_class and test_novel_class `. Note that both the support set and query set in the `valset` are different from the ones in the `testset`. File Structure and Content (All files are in `.tif` format): ----------------------------------------------------------- 1. **trainset.zip**: - Contains `images` and `labels` folders - `images` folder: 258 images of size 1024x1024 with a GSD (Ground Sampling Distance) of 0.6-1m. - `labels` folder: 258 segmentation masks of the images in the `images` folder. 2. **valset.zip**: - Contains `images` and `labels` folders - `images` folder: 50 images of size 1024x1024 with a GSD (Ground Sampling Distance) of 0.6-1m. - `labels` folder: 20 labels of the ``support set`` images in the `images` folder. The labels for the 30 ``query set`` images in the `images` folder are withheld.3. **testset.zip**: - Contains `images` and `labels` folders - `images` folder: 100 images of size 1024x1024 with a GSD (Ground Sampling Distance) of 0.6-1m. - `labels` folder: 20 labels of the ``support set`` images in the `images` folder. The labels for the 80 ``query set`` images in the `images` folder are withheld. 4. **train.txt**: - Contains a list of file names in the `trainset.zip`. 3. **val.json** and **test.json**: - Contains a list of file names the in the `valset.zip` and `testset.zip`, respectively. Below is the structure of the `val.json` and `test.json` files. - fnames = { {"support_set": {8: ["filename_1.tif", "filename_2.tif", ...., "filename_5.tif"], 9: ["filename_1.tif", "filename_2.tif", ...., "filename_5.tif"], 10: ["filename_1.tif", "filename_2.tif", ...., "filename_5.tif"], 11: ["filename_1.tif", "filename_2.tif", ...., "filename_5.tif"]}, {"query_set": ["filename_1.tif", "filename_2.tif", "filename_3.tif", ... ...., "filename_n.tif"]} } Land Cover Mapping Classes Strucure: ------------------------------------1. **The `trainset`: classId2className = { # ***Base classes*** 1: 'tree', 2: 'rangeland', 3: 'bareland', 4: 'agric land type 1', 5: 'road type 1', 6: 'sea, lake, & pond', 7: 'building type 1' }2. **The `valset` and `testset`: classId2className = { # ***Base classes*** 1: 'tree', 2: 'rangeland', 3: 'bareland', 4: 'agric land type 1', 5: 'road type 1', 6: 'sea, lake, & pond', 7: 'building type 1' # ***Novel classes*** 8: '', 9: '', 10: '', 11: '' } - The class names for the ***Novel classes*** depends on the data set. For the `valset`, the class names can be updated as: { 8: 'road type 2', 9: 'river', 10: 'boat & ship', 11: 'agric land type 2' } For the `testset`, the class names can be updated as: { 8: 'vehicle & cargo-trailer', 9: 'parking space', 10: 'sports field', 11: 'building type 2' }License See OpenEarthMap
本次挑战赛已结束，请使用本版本开展赛后研究。本版本包含全部数据与文件，但测试集查询集(query-set)的标签已隐藏，研究人员可通过挑战赛排行榜提交预测结果以进行评估。 ### 概述 本挑战赛与L3D-IVU 2024 CVPR workshop联合举办，是面向广义少样本语义分割(Generalized Few-Shot Semantic Segmentation, GFSS)任务的OpenEarthMap基准数据集的扩展挑战赛。本挑战赛旨在针对OpenEarthMap数据集上的少样本语义分割学习方法开展评估与基准测试，以推动服务于社会公益的地理信息科学研究。其核心动机是助力研究人员开发适用于高分辨率遥感(Remote Sensing, RS)图像语义分割的少样本学习算法。 ### 官方页面 https://cliffbb.github.io/OEM-Fewshot-Challenge/ ### 基线模型 本挑战赛的基线模型可在此处获取。 ### 排行榜 挑战赛排行榜已开放用于赛后研究，支持研究人员针对测试集查询集的预测结果进行评估。 ### 数据集说明 本数据集专为遥感少样本学习设计，尤其适用于土地覆盖制图中的广义少样本语义分割任务。数据集从原始OpenEarthMap遥感图像语义分割基准数据集中选取了408个样本。该数据集将原始OpenEarthMap基准的8个语义类别扩展至15个，并分别按7:4:4的比例划分为互不相交的训练基类(train_base_class)、验证新类(val_novel_class)与测试新类(test_novel_class)集合（即train_base_class ∩ val_novel_class ∩ test_novel_class = ∅）。 408个样本进一步划分为训练集(trainset，258个)、验证集(valset，50个)与测试集(testset，100个)。其中训练集用于骨干网络的预训练，仅包含训练基类划分下的图像与标签。验证集与测试集均包含支持集(support set)与查询集(query-set)，对应5样本少样本场景，涉及4个新类与7个基类的广义少样本语义分割任务。验证集与测试集分别包含验证新类与测试新类划分下的图像与标签。 本挑战赛分为两个阶段：开发阶段与评估阶段。验证集用于开发阶段，测试集用于评估阶段。验证集与测试集均包含20组图像-标签对：支持集内为4个新类各5组样本。验证集与测试集的查询集分别额外包含30张与80张图像，需通过各自支持集内的20张带标签图像完成预测。 支持集内每张图像的标签均不包含任何训练基类划分下的类别。此外，每组5样本样本中，标签仅对应一个新类（即每组5样本样本仅对应一个新类）。值得注意的是，验证集查询集图像的标签同时包含训练基类与验证新类；而测试集查询集图像的标签同时包含训练基类与测试新类。需特别说明：验证集与测试集的支持集和查询集均互不相同。 ### 文件结构与内容（所有文件均为.tif格式） ----------------------------------------------------------- 1. **trainset.zip**： - 包含`images`与`labels`文件夹 - `images`文件夹：258张分辨率为1024×1024的图像，地面采样距离(Ground Sampling Distance, GSD)为0.6-1米。 - `labels`文件夹：对应`images`文件夹内258张图像的分割掩码。 2. **valset.zip**： - 包含`images`与`labels`文件夹 - `images`文件夹：50张分辨率为1024×1024的图像，地面采样距离(GSD)为0.6-1米。 - `labels`文件夹：包含`images`文件夹内支持集图像的20组标签。`images`文件夹内30张查询集图像的标签已隐藏。 3. **testset.zip**： - 包含`images`与`labels`文件夹 - `images`文件夹：100张分辨率为1024×1024的图像，地面采样距离(GSD)为0.6-1米。 - `labels`文件夹：包含`images`文件夹内支持集图像的20组标签。`images`文件夹内80张查询集图像的标签已隐藏。 4. **train.txt**： - 包含`trainset.zip`内的文件名列表。 5. **val.json**与**test.json**： - 分别包含`valset.zip`与`testset.zip`内的文件名列表。以下为`val.json`与`test.json`的文件结构： fnames = { "support_set": { 8: ["filename_1.tif", "filename_2.tif", ..., "filename_5.tif"], 9: ["filename_1.tif", "filename_2.tif", ..., "filename_5.tif"], 10: ["filename_1.tif", "filename_2.tif", ..., "filename_5.tif"], 11: ["filename_1.tif", "filename_2.tif", ..., "filename_5.tif"] }, "query_set": ["filename_1.tif", "filename_2.tif", "filename_3.tif", ..., "filename_n.tif"] } ### 土地覆盖制图类别结构 ------------------------------------ 1. **训练集(trainset)** classId2className = { # ***基类(Base classes)*** 1: '树木', 2: '草原', 3: '裸地', 4: '一类农业用地', 5: '一类道路', 6: '海洋、湖泊与池塘', 7: '一类建筑' } 2. **验证集(valset)与测试集(testset)** classId2className = { # ***基类(Base classes)*** 1: '树木', 2: '草原', 3: '裸地', 4: '一类农业用地', 5: '一类道路', 6: '海洋、湖泊与池塘', 7: '一类建筑', # ***新类(Novel classes)*** 8: '', 9: '', 10: '', 11: '' } - 新类的类别名称取决于数据集具体配置： 针对验证集，新类名称可更新为： { 8: '二类道路', 9: '河流', 10: '船舶', 11: '二类农业用地' } 针对测试集，新类名称可更新为： { 8: '车辆与货运挂车', 9: '停车场', 10: '运动场', 11: '二类建筑' } ### 许可证 详见OpenEarthMap
