Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Continual Gaussian Mixture Distribution Modeling for Class Incremental Semantic Segmentation
Authors: Guilin Zhu, Runmin Wang, Yuanjie Shao, Wei dong Yang, Nong Sang, Changxin Gao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Pascal VOC and ADE20K show that our method achieves superior performance compared to previous methods, especially in more challenging long-term incremental scenarios. The paper contains a dedicated section 4 titled 'Experiments', detailing quantitative comparisons and ablation studies on public datasets. |
| Researcher Affiliation | Academia | 1National Key Laboratory of Multispectral Information Intelligent Processing Technology, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, 2School of Information Science and Engineering, Hunan Normal University EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Dynamic Adjustment Strategy Input: Training dataset Dt+1, extractors f t and f t+1, stored GMMs {ϕ c}C1:t c=1, training interval e Output: Updated GMMs {ˆϕ c}C1:t c=1 1 while the end of each interval e do 2 Freeze extractors f t and f t+1; 3 for (xt, yt) Dt+1 do 4 Compute the mask maps m based on Eq. 6; 5 Obtain old class features F c from extractor f t+1 based on Eq. 7; 7 Initialize GMMs with {ϕ c}C1:t c=1; 8 for c in C1:t do 9 Sample pseudo-features Fc = S(ϕ c, Nc); 10 while not converged do 11 Estimate responsibility γk = πck N ({Fc,F c}|µck,Σck) PK j=1 πcj N ({Fc,F c}|µcj,Σcj) ; /* The E-step 12 Update the parameters {πc, µc, Σc} by Eq. 4; ; /* The M-step 14 ˆϕ c {πc, µc, Σc}; |
| Open Source Code | Yes | The source code is available at https://github.com/zhu-gl-ux/Co Ga Mi D |
| Open Datasets | Yes | We conduct comprehensive experiments on two public datasets: Pascal VOC 2012 [15] and ADE20K [57]. |
| Dataset Splits | Yes | We follow the protocols in [4] to evaluate our model across various incremental scenarios defined as as Nb Nn, where Nb and Nn denote numbers of base and novel classes, respectively. For instance, in 15-1 scenario, training begins with 15 classes, followed by the addition of one new class at each incremental step. Pascal VOC 2012 contains 10,582 training images and 1449 validation images, encompassing 20 foreground classes. ADE20K includes 20,210 training images and 2,000 validation images, distributed across 150 classes. |
| Hardware Specification | Yes | We conduct experiments on four NVIDIA RTX 4090 GPUs using PyTorch. |
| Software Dependencies | No | The paper mentions 'PyTorch' but does not provide specific version numbers for it or any other software components. |
| Experiment Setup | Yes | We use SGD optimizer to optimize the network. The learning rate for the initial step is set to 0.001 and 0.00025 for Pascal VOC 2012 and ADE20K, respectively, and is reduced by a factor of 0.1 for the incremental steps. We train the network for 60 epochs on Pascal VOC and 100 epochs on ADE20K with 0.9 momentum and 0.0001 weight decay in all steps. The batch size is set to 24 for both datasets. For the hyper-parameters , α, β and K are set to 5, 0.05, and 3, respectively. |