LIS: Low-light Instance Segmentation Dataset (IJCV 2023)

This is the official release of the LIS dataset from the IJCV 2023 paper "Instance Segmentation in the Dark" (Chen, Fu, Wei, Zheng, Heide). It contains 2,051 paired low-light / normal-light images with instance-level pixel-wise annotations, captured by a Canon EOS 5D Mark IV under very low illumination (0.01–1 lux).

本数据集为 IJCV 2023 论文《Instance Segmentation in the Dark》配套的 LIS 数据集官方发布版，包含 2,051 对低光/正常光图像及实例级像素标注，由佳能 EOS 5D Mark IV 在 0.01–1 lux 极低照度下拍摄。

Quick Start / 快速上手

One command to download / 一条命令下载：

pip install -U huggingface_hub
huggingface-cli download Linwei-Chen/LIS --repo-type dataset --local-dir LIS

Use it with the official code / 配合官方代码使用（下载得到的目录就是代码期望的 data/LIS/ 布局，无需重新整理 / the downloaded folder is exactly the data/LIS/ layout the code expects):

git clone https://github.com/Linwei-Chen/LIS && cd LIS
huggingface-cli download Linwei-Chen/LIS --repo-type dataset --local-dir data/LIS
huggingface-cli download Linwei-Chen/LIS-models \
    mask_rcnn_r50_fpn_caffe_AWD_SCB_DSL_SynCOCO2LIS=31.8.pth --local-dir checkpoints
ln -sfn ../data mmdetection/data && cd mmdetection
python tools/test.py ../Configs/mask_rcnn_r50_fpn_caffe_AWD_SCB_DSL_SynCOCO2LIS.py \
    ../checkpoints/mask_rcnn_r50_fpn_caffe_AWD_SCB_DSL_SynCOCO2LIS=31.8.pth --eval bbox segm
# → segm mAP 32.0 / bbox mAP 37.5 on this privacy-filtered release (measured)
#   (paper: 31.8 / 37.6 on the full 669-image test set)

Training, single-image inference, and the plug-and-play usage are documented in the code repo README. / 训练、单图推理与即插即用用法见代码仓库 README。

Dataset Structure / 数据结构

LIS/
├── RGB_Normal/JPEGImages/    # normal-light sRGB images (1200×800 png, odd ids)
├── RGB_Dark/JPEGImages/      # low-light camera original JPGs (even ids)
├── RAW_Normal/JPEGImages/    # normal-light images from high bit-depth RAW (png)
├── RAW_Dark/JPEGImages/      # low-light images from RAW, "NoRatio" variant (png)
├── annotations/              # COCO-format json files (see below)
├── EXCLUDED_IDS.txt
└── RELEASE_STATS.txt

• Pairing rule / 配对规则: every normal-light image (odd id N) has a low-light counterpart (even id N+1). / 每张正常光图（奇数编号 N）对应一张低光图（偶数编号 N+1）。
• The four folders share the same 2,051 scenes. / 四个文件夹对应同样的 2,051 个场景。
• Naming follows the original drive release / 命名与原始网盘发布一致: RGB_Normal/RAW_Dark are byte-identical to the original release; RGB_Dark and RAW_Normal are the original release renditions (the full-quality working versions on the author's machine differ only in JPG/PNG encoding). / RGB_Normal/RAW_Dark 与原始发布逐字节一致；RGB_Dark 与 RAW_Normal 即原始发布版本（作者机器上的全质量工作版仅编码差异）。
• RAW_Normal/RAW_Dark are rendered from 14-bit RAW dumps; RAW_Dark is the NoRatio variant (no normalization ratio applied) used by the paper's RAW-input protocol. / RAW_Normal/RAW_Dark 由 14-bit RAW 解包渲染；RAW_Dark 为论文 RAW 输入协议使用的 NoRatio 版本。

Example images / 示例图像

Dark / normal pairs (canteen scene id 1003, and id 1013) / 暗-亮成对图像（食堂场景 id 1003 与 id 1013）：

 

Instance annotations with true class names / 按真实类名渲染的实例标注（左：id 1003 normal 侧 14 实例，chair / diningtable；右：id 169 dark 侧 3 实例，bus）：

 

Annotations / 标注

COCO format, 8 categories. True label semantics (category id → visual class, 0-based): 0=bicycle, 1=car, 2=motorbike, 3=bus, 4=bottle, 5=chair, 6=diningtable, 7=tvmonitor. ⚠️ The categories/name block inside the json files is historically misaligned with these true semantics — read the Category mapping section below before writing any class-dependent code. / COCO 格式，8 个类别。真实类别语义（类别 id → 视觉类别，0 起始）：0=自行车, 1=汽车, 2=摩托车, 3=公交, 4=瓶子, 5=椅子, 6=餐桌, 7=显示器。⚠️ json 文件内的 categories 名字块与真实语义存在历史错位——写任何依赖类名的代码前，务必先阅读下方 Category mapping（类别映射） 一节。

• lis_coco_png_*: annotations for the png protocol (RGB_Normal / RAW_Normal / RAW_Dark), file_name = ./JPEGImages/N.png. Train: 1,447 images / 6,999 instances; test: 604 images / 2,680 instances. / png 协议标注，训练 1,447 图 / 6,999 实例，测试 604 图 / 2,680 实例。
• The non-+1 lis_coco_JPG_* files reference the normal-light side and are kept as in the original release (which shipped them without images); if images are needed, use the RGB_Normal/ pngs by id. / 不带 +1 的 lis_coco_JPG_* 引用正常光侧，按原始发布原样保留（原始发布亦未附带这批图）；如需图片可按编号使用 RGB_Normal/ 的 png。
• lis_coco_JPG_*+1: annotations for the camera-JPG dark-side protocol (RGB_Dark), file_name = ./JPEGImages/N.JPG. / 相机 JPG 协议标注。
• coco.json / coco+1.json: earlier annotation versions kept for completeness. / 更早期的标注版本，供完整参考。
• *_x_* files (10x/20x/30x/40x/50x/100x): gain-subset splits. / 增益倍率子集划分。

Category mapping (important) / 类别映射（重要）

True label semantics / 真实类别语义. All 97 lis_coco_* annotation files use 0-based category ids. The true visual class of each id, verified instance-by-instance on multiple scenes, is:

所有 97 个 lis_coco_* 标注使用 0 起始的类别 id。每个 id 对应的真实视觉类别已逐类目击验证：

id	0	1	2	3	4	5	6	7
true class / 真实类别	bicycle	car	motorbike	bus	bottle	chair	diningtable	tvmonitor

The categories/name block inside the json files is historically misaligned with these true semantics (it claims 0=bicycle, 1=chair, 2=diningtable, 3=bottle, 4=motorbike, 5=car, 6=tvmonitor, 7=bus — but e.g. the id-1 instances are visually cars, not chairs). The json files themselves are not going to be renamed, for the reason below. / json 文件内的 categories 名字块与上述真实语义系统性错位（名字块声称 0=bicycle, 1=chair, 2=diningtable…，但 id 1 的实例视觉上是汽车而非椅子）。由于下述原因，发布 json 保持原样、不做改名。

Why the published numbers are unaffected / 为何已发布数字不受影响. mmdetection binds label ids to class names via a name→id lookup (get_cat_ids(cat_names=...)), and the released configs' class tuple ('bicycle','car','motorbike','bus','bottle','chair','diningtable','tvmonitor') follows exactly the true id order above. The resulting training/evaluation label space is therefore self-consistent, and all paper numbers and released checkpoints reproduce as published. / mmdetection 通过「名字→id 查表」把类别名绑定到标签 id，而发布 config 的类别元组恰好就是上表的真实 id 序，因此训练/评测的标签空间自洽——论文全部数字与发布权重均可照常复现。

Do NOT rename the json categories or reorder the config class tuple — either change breaks compatibility with the released checkpoints. If you train from scratch with these jsons you may either keep everything as released (recommended), or fix the name block to the true semantics and use the true-order tuple; label-permutation invariance guarantees identical mAP. / 不要改名 json 的类别名，也不要调整 config 的类别元组顺序——任一改动都会破坏与已发布权重的评测兼容性。若用本数据从头训练，可保持原样（推荐），也可同时修正名字块+元组为真实语义（标签置换不变性保证 mAP 完全相同）。

Per-class instance counts with true names (this filtered release) / 分类别实例数（真实类名，本过滤发布版）:

class	train inst	train imgs	test inst	test imgs
bicycle	849	251	421	158
car	2,503	684	833	280
motorbike	399	185	203	106
bus	769	237	118	42
bottle	1,031	138	291	42
chair	950	270	503	124
diningtable	339	192	240	115
tvmonitor	159	85	71	40
total	6,999	1,447 images	2,680	604 images

(One image may contain several classes, so the per-class image columns sum to more than the file counts. / 一张图可含多个类别，故分类别图数列之和大于文件总数。)

coco.json / coco+1.json are early drafts: they use 1-based ids with a different category order (1=car, 2=motorbike, 3=bicycle, 4=chair, 5=diningtable, 6=bottle, 7=tvmonitor, 8=bus), a different file_name format, and contain a few out-of-bounds boxes. Do not mix them with the lis_coco_* files. / coco.json / coco+1.json 为早期草稿：使用 1 起始编号与另一套类别顺序（1=car, 2=motorbike, …, 8=bus）、file_name 格式不同、且含少量越界框，不可与 lis_coco_* 混用。

In the +1 annotation files, image_id keeps the normal-side odd numbering (so both sides of a pair share the same id), while file_name points to the dark-side image (id + 1). This offset is by design. / 带 +1 的标注文件中，image_id 保持正常光侧奇数编号（成对图共用同一 id），file_name 指向低光侧图片（编号 +1）——此为设计约定，不是错位。

LIS 1.1 revised annotations / 1.1 修订标注

The authors have released a revised annotation set (LIS 1.1, ~100 images / ~312 instances with improved boxes and masks) via issue #23 (Baidu Drive, code: IJCV). Note that LIS 1.1 uses 1-based category ids in a different order — the same order as the early drafts (1=car, 2=motorbike, 3=bicycle, 4=chair, 5=diningtable, 6=bottle, 7=tvmonitor, 8=bus) — so always read its own categories block, and do not reuse the lis_coco_* id semantics or the +1 dark-side pairing convention with it. / 作者已通过 issue #23（百度网盘，提取码 IJCV）发布修订版标注（LIS 1.1，约 100 图 / 312 实例，框与掩码更准）。注意 LIS 1.1 使用 1 起始编号与另一套类别顺序（与早期草稿相同：1=car … 8=bus），使用前以其自带 categories 块为准；不要把 lis_coco_* 的 id 语义或 +1 暗侧配对约定套用到 1.1 上。

Known version differences / 标注版本差异说明

The full unfiltered png-protocol annotation contains 10,389 instances over 2,230 images; the paper's text reports 10,504, which corresponds to the JPG-protocol version. In this (privacy-filtered) release, lis_coco_png_traintest.json contains 9,679 instances, and the early-draft coco+1.json contains 9,052 (9,733 before filtering). These are annotation versions from different release stages, all provided here. / 全量未过滤 png 协议标注 10,389 实例（2,230 图）；论文正文 10,504 对应 JPG 协议版本。本隐私过滤版发布中 lis_coco_png_traintest.json 为 9,679 实例，早期草稿 coco+1.json 为 9,052（过滤前 9,733）。以上为不同阶段的标注版本，均一并提供。

全量未过滤标注 png 协议共 10,389 实例（2,230 图）；论文正文所写 10,504 对应 JPG 协议口径。经隐私过滤后，发布的 lis_coco_png_traintest.json 为 9,679 实例。更早版本（coco+1.json）为 9,052（过滤前 9,733）。以上为不同阶段的标注版本，均一并提供。

Privacy Filtering / 隐私过滤

The raw capture contains 2,230 pairs; 179 pairs are excluded from this public release (listed in EXCLUDED_IDS.txt as normal-side odd ids; the dark side is id+1):

原始采集共 2,230 对，本公开版排除 179 对（清单见 EXCLUDED_IDS.txt，记录 normal 侧奇数编号，对应低光侧为其 +1）：

1. Close-up shots with readable license plates. / 近景中车牌可辨认的图像。
2. Images with large bright screens whose content was not reviewed in full, excluded out of caution. / 含大面积亮屏的图像，因屏幕内容未逐张复核而保守排除。

Because the paper's evaluation runs on the full test set (669 images), results reported in the paper use 2934 test instances, while the released test annotation has 2,680 (604 images); numbers may differ slightly when re-evaluating on the filtered release (measured, R50 Ours: segm 32.0 / box 37.5 on the release vs 31.8 / 37.6 on the full set). The full-resolution camera-original JPGs of excluded images are never published.

论文评测在完整测试集（669 图）上进行，而发布版测试集为 604 图，复现时数字可能存在细微差异（实测：R50 Ours 在发布版上 segm 32.0 / bbox 37.5，完整测试集上为 31.8 / 37.6）。被排除图像的相机原档不做公开发布。

License / 许可

The LIS dataset is released under CC BY 4.0.

LIS 数据集以 CC BY 4.0 许可发布。

Citation / 引用

If you use this dataset, please cite the paper and the previous LOD dataset work:

使用本数据集请引用论文及前作 LOD 数据集：

@article{2023lis,
  title={Instance Segmentation in the Dark},
  author={Chen, Linwei and Fu, Ying and Wei, Kaixuan and Zheng, Dezhi and Heide, Felix},
  journal={International Journal of Computer Vision},
  volume={131},
  number={8},
  pages={2198--2218},
  year={2023},
  publisher={Springer}
}

@inproceedings{Hong2021Crafting,
  title={Crafting Object Detection in Very Low Light},
  author={Yang Hong, Kaixuan Wei, Linwei Chen, Ying Fu},
  booktitle={BMVC},
  year={2021}
}

Code & models: github.com/Linwei-Chen/LIS · Linwei-Chen/LIS-models