Low-light video processing remains an underexplored problem in computer vision, primarily due to the scarcity of high-quality ground-truth data. Existing public datasets exhibit several limitations, including restricted motion diversity, limited scene coverage, weak degradations, reliance on auxiliary modalities such as event data, and misalignment between low-light inputs and ground truth. Furthermore, current datasets typically address either low-light enhancement or scene understanding, whereas these tasks should be considered jointly. We introduce OLiVES, a benchmark dataset for outdoor low-light videos with dynamic scenes, designed to support both low-light video enhancement (LLVE) and video object segmentation (VOS). The dataset addresses key content gaps by incorporating a diverse range of natural textures, animals, occlusions, dynamic activities, and close-up scenes. To the best of our knowledge, OLiVES is the first outdoor low-light video dataset captured using professional cinema-grade cameras. It provides over 25,000 aligned normal- and low-light frame pairs, along with over 46,000 VOS annotations. We benchmark state-of-the-art LLVE and VOS methods on OLiVES and show that training on it improves performance on both tasks. The OLiVES project page is available at: https://anonymous-OLiVES.github.io/.
Figures & tables
Figure 1: Examples of the OLiVES dataset. (a) Low-light videos with their corresponding normal-light ground truth and object segmentation masks. Each scene is provided at multiple light levels, (b) illustrating the noise levels exhibited in the low-light videos when using different ND filters (histogram-matched to GT for comparison).
Dataset
Modality
Frames
Cam. Motion
Obj. Motion
Resolution
VOS ann.
Device
DRV [ 3 ]
Raw
22,220
Static
Static
5496×3672
✗
Sony RX100 VI
SDSD [ 40 ]
RGB
37,500
Dynamic
Static
1920×1080
✗
Canon 6D Mark II
LLRVD [ 14 ]
Raw
4,086
Dynamic
Dynamic
2600×1400
✗
Sony α 7R IV
BVI-RLV [ 31 ]
RGB
31,800
Dynamic
Static
1920×1080
✗
Sony α 7S II
SDE [ 4 ]
RGB/Event
31,477
Dynamic
Static
346×260
✗
DAVIS 346
DID [ 13 ]
RGB
41,038
Dynamic
Static
2560×1440
✗
5 Camera Models
Table 1: Comparison of low-light video enhancement datasets.
Dataset
Light
Frames
Categories
Objects
Annotations
New objects
DAVIS 2016 [ 35 ]
Normal
3,440
-
50
3,455
✗
DAVIS 2017 [ 36 ]
Normal
6,208
-
205
13,543
✗
YouTube-VOS [ 42 ]
Normal
120,532
94
7755
197,272
✓
MOSEv1 [ 10 ]
Normal
130,149
36
5,200
431,725
✗
MOSEv2 [ 11 ]
Mixed
468,251
200
10,074
701,976
✗
LLE-VOS [ 23 ]
Low
5,600
-
120
9,606
✗
Table 2: Comparison of video object segmentation (VOS) datasets.
Figure 2: Capture setup. (a) Our beam-splitter rig with two genlocked cinema cameras mounted on a shared tripod. (b) Schematic of the optical path: incoming light is split between Camera 1, which records the normal-light ground truth Igt , and Camera 2, which records the low-light frame Ilow through one of three ND filters (1/16, 1/32, 1/64). Genlock ensures temporal synchronization between the two streams.
Figure 3: Registration pipeline of the OLiVES dataset.
Figure 4: Finetuning the fine feature encoder of RoMa with synthetically degraded low-light pairs and a contrastive objective to improve robustness to low-light degradations.
Table 3: Summary of OLiVES VOS benchmark for the train, validation and test set.
Method
Venue
Type
Data
PSNR ↑
SSIM ↑
LPIPS ↓
tLPIPS ↓×10−2
WE ↓×10−2
VMAF ↑
SNR-aware [ 43 ]
CVPR’22
S
I
26.20
0.835
0.268
6.47
5.97
73.51
LLFlow [ 41 ]
AAAI’22
S
I
21.65
0.862
0.219
2.82
3.81
69.50
Retinexformer [ 2 ]
ICCV’23
S
I
27.65
0.834
0.241
6.35
5.41
76.93
LightenDiffusion [ 21 ]
ECCV’24
UP
I
20.32
0.741
0.380
8.70
5.56
54.36
Wakeup-darkness [ 51 ]
ACM TOMM’25
US
I
17.46
0.534
0.630
8.82
9.26
71.39
SGZSL [ 52 ]
WACV’22
US
V
16.61
0.527
0.693
10.70
11.80
74.74
Table 4: Benchmarking on OLiVES dataset across recent LLVE and LLIE SOTA methods. Bold and underline denote the best and second-best performances among all methods respectively.
Method
Venue
OLiVES val
OLiVES test
LLE-VOS [ 23 ] val
J&F
J
F
B
J&F
J
F
B
J&F
J
F
STCN [ 8 ]
NeurIPS’21
89.4
87.2
91.5
83.5
78.3
73.8
82.9
70.8
32.3
37.9
26.7
AOT [ 47 ]
NeurIPS’21
91.5
88.5
94.6
85.1
80.5
75.8
85.1
73.4
54.1
57.5
50.7
De-AOT [ 48 ]
NeurIPS’22
91.5
88.4
94.5
85.0
86.9
82.0
91.7
80.0
62.6
65.6
59.7
XMem [ 7 ]
ECCV’22
86.9
84.4
89.5
79.5
74.4
68.9
79.8
65.3
53.0
56.0
50.1
Cutie [ 6 ]
CVPR’24
93.1
91.3
94.8
88.0
88.3
84.2
92.4
82.0
59.2
62.8
55.7
Table 5: Benchmark of SOTA VOS methods trained on the OLiVES dataset and evaluated on OLiVES validation and test set, and LLE-VOS validation set. Bold and underline denote the best and second-best performances among all methods respectively.
LLVE Method
LLE-VOS [ 23 ] val
J&F
J
F
Cutie (off-the-shelf) [ 6 ]
53.2
58.5
48.0
LLVE-STCD [ 45 ] + Cutie
57.9
62.1
53.8
BVI-Mamba [ 20 ] + Cutie
58.1
62.0
53.8
RetinexFormer [ 2 ] + Cutie
58.3
61.9
54.6
Table 6: Quantitative results of two-stage methods on LLE-VOS validation set, using LLVE methods trained on OLiVES and pre-trained Cutie [ 6 ] .
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Method
AUC@5
AUC@10
AUC@20
Pretrained
62.28
76.45
86.25
Replace Coarse Features
62.19
76.35
86.15
Replace Fine Features
52.33
68.68
80.79
Appendix
Table 7: Ablation study on the impact of coarse and fine features in RoMa. We replace the corresponding features in the pretrained model with those extracted from degraded inputs.
Method
Non-degraded
Degraded
AUC@5
AUC@10
AUC@20
AUC@5
AUC@10
AUC@20
Pretrained
62.28
76.45
86.25
54.75
70.42
82.04
I: Deg. Aug.
+0.34
+0.11
+0.04
+2.43
+2.06
+1.46
II: I+VICReg
+0.10
-0.03
-0.06
+2.74
+2.17
+1.49
III: II+InfoNCE
+0.56
+0.38
+0.17
+3.12
+2.44
+1.66
Appendix
Table 8: Finetuning performance of RoMa on non-degraded and degraded MegaDepth-1500. All values (except the first row) indicate absolute improvements over the pretrained baseline.
Figure 6: Qualitative results of LLVE methods trained on OLiVES.
Figure 7: Custom annotation software to create manual annotations for OLiVES VOS benchmark.
Figure 8: Qualitative comparison of Cutie on an example video from MOSEv2 [ 11 ] (reduced frame brightness for better segmentation visibility). (Left) the first frame with the input masks. (Right) Crops of segmentation outputs for Cutie [ 6 ] with various training data.
Size
Count
OLiVES test
J&F
J
F
Small
5
-
-
-
Medium
40
-
-
-
Large
109
-
-
-
Appendix
Table 9: Quantitative results on OLiVES test set, grouped by size of the objects. Count is the total count of the object sizes.
Figure 9: Examples of difficult segmentation scenarios in OLiVES.
Video instance segmentation (VIS) for low-light content remains highly challenging for both humans and machines alike, due to noise, blur and other adverse conditions. The lack of large-scale annotated datasets and the limitations of current synthetic pipelines, particularly in modeling temporal degradations, further hinder progress. Moreover, existing VIS methods are not robust to the degradations found in low-light videos and, consequently, perform poorly even after finetuning. In this paper, we introduce ELVIS (Enhance Low-Light for Video Instance Segmentation), a framework that enables domain adaptation of state-of-the-art VIS models to low-light scenarios. ELVIS is comprised of an unsupervised synthetic low-light video pipeline that models both spatial and temporal degradations, a calibration-free degradation profile estimation network (VDP-Net) and an enhancement decoder head that disentangles degradations from content features. ELVIS improves performances by up to +3.7AP on the synthetic low-light YouTube-VIS 2019 dataset and beats two-stage baselines by at least +2.8AP on real low-light videos. Code and dataset available at: https://joannelin168.github.io/research/ELVIS
Joanne Lin, Ruirui Lin, Yini Li +2
Visual Information Laboratory, University of Bristol
Accurate text recognition in low-light environments is essential for intelligent systems in applications ranging from autonomous vehicles to smart surveillance. However, challenges such as poor illumination and noise interference remain underexplored. To address this gap, we introduce LSTR, a large-scale Low-light Scene Text Recognition dataset comprising 11,273 low-light images generated from well-lit datasets (ICDAR2015, IIIT5K, and WordArt), along with ESTR, which includes 60 real nighttime street-scene images in English and Spanish for exclusive evaluation. We explore two solution strategies: (1) employing Optical Character Recognition (OCR) models with fine-tuning and LoRA-based fine-tuning and (2) a joint training strategy that integrates a low-light image enhancement (LLIE) module with an OCR model. In particular, we propose a novel re-render LLIE (RLLIE) module, which demonstrates improved performance on real-world data. Through extensive experimentation, we analyze various training strategies and address a key research question: How bright is bright enough for effective scene text recognition? Our results indicate that standalone LLIE or OCR models perform inadequately under low-light conditions, highlighting the advantages of specialized, jointly trained text-centric approaches. Additionally, we provide a comprehensive benchmark to support future research in robust low-light scene text recognition. https://huggingface.co/datasets/lumimusta/Low-light_Scene_Text_Dataset.
This paper presents a review of the NTIRE 2026 Low-light Enhancement: Twilight Cowboy Challenge. The objective of the competition was to merge a set of misaligned smartphone images in the raw domain, captured in low-light conditions, into a single, clean image. Introduced setup simultaneously addresses two problems of low-light photography: visual degradations such as high noise and mixed scene illuminants, and the geometric inconsistencies caused by hand movement during multi-frame capture. To advance research in low-light and nighttime computational photography, a challenging dataset was collected comprising 585 real-world scenes, spanning indoor low-light and outdoor nighttime conditions, for training and benchmarking participant solutions. The competition employed a three-stage evaluation protocol: automatic validation via the CodaBench platform in stages one and two, followed by blind assessment on a private test set for the final ranking. Ten teams surpassed the established baseline, achieving improvements of up to +6.49 dB in PSNR and +0.0101 in SSIM, thereby establishing new state-of-the-art performance for burst-based low-light image enhancement. These results demonstrate significant progress in handling real-world noise, motion, and illumination variability in the low-light setting. Comprehensive results, leaderboards, and additional information are publicly available at https://nightimaging.org.