Coming Out of the Dark: Human Pose Estimation in Low-light Conditions
Authors: Yong Su, Defang Chen, Meng Xing, Changjae Oh, Xuewei Liu, Jieyang Li
IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on low-light images validate the effectiveness of our approach, with the LLIP dataset and MHFCL being instrumental to its success. Furthermore, MHFCL demonstrates competitive performance in HPE tasks, underscoring its robustness and versatility. |
| Researcher Affiliation | Collaboration | Yong Su1 , Defang Chen1 , Meng Xing2,3 , Changjae Oh4 , Xuewei Liu5 and Jieyang Li1 1Tianjin Normal University 2University of Science and Technology of China 3Ningbo Institute of Digital Twin, Eastern Institute of Technology 4Queen Mary University of London 5Research & Development Branch, FAW Toyota Motor Co.,Ltd. |
| Pseudocode | No | The paper describes the proposed framework MHFCL, its RHR and VPE streams, and the Multi-grained Feature Consistency Learning mechanism using textual descriptions and an architectural diagram (Figure 4). However, no explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | No | The paper states: "The dataset is available on the project website: https://llip2024.github.io". This explicitly refers to the dataset and not the source code for the proposed methodology (MHFCL). There is no other statement or link indicating the release of the code for MHFCL. |
| Open Datasets | Yes | To alleviate the issue, we construct a Low-Light Images and Poses (LLIP) dataset, which includes only paired low-light images and pose annotations obtained using off-the-shelf motion capture devices. [...] The dataset is available on the project website: https://llip2024.github.io. [...] We conduct pose estimation experiments using three datasets: the proposed LLIP dataset, [...] and the widely used MS-COCO keypoint dataset [Lin et al., 2014]. |
| Dataset Splits | Yes | The LLIP dataset consists of 12,378 low-light images for training and 4,814 images for testing. The training set includes images of eight individuals in two different scenes, while the test set spans five distinct scenes, allowing for an evaluation of model generalizability across varied conditions. [...] Figure 3 illustrates the distribution of training and testing samples across these categories, offering a detailed overview of the dataset composition. Notably, samples in the ELL and FLL categories account for over 70% of the testing set, making the LLIP dataset particularly challenging for low-light pose estimation tasks. |
| Hardware Specification | Yes | Our experiments were conducted on an RTX 3090 platform with a batch size of 6, utilizing the Adam optimizer. |
| Software Dependencies | No | The paper mentions "Adam optimizer" but does not specify any other software dependencies like programming languages, libraries, or frameworks with their version numbers required to replicate the experiments. |
| Experiment Setup | Yes | Our experiments were conducted on an RTX 3090 platform with a batch size of 6, utilizing the Adam optimizer. During the pre-training phase of the RHR stream, we augmented the LOL training set with an additional 303 images from the LLIP dataset, resulting in a total of 788 images. The initial learning rate was set to 1 × 10−4. For the VPE stream, the learning rate was set to 2 × 10−4. Both streams were pretrained for 40 epochs. [...] Joint training was then performed for 100 epochs. During the pretraining phase of the RHR stream, we applied regularization coefficients: λ1 = 0, λ2 = 0.1, λ3 = 1, and λ4 = 0.5. For the pre-training phase of the pose estimation network, we set λ1 = 1 and λ2 = λ3 = λ4 = 0. During the joint fine-tuning phase, regularization coefficients were adjusted to λ1 = 1, λ2 = λ3 = λ4 = 0.1. |