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..
Searching Efficient Semantic Segmentation Architectures via Dynamic Path Selection
Authors: Yuxi Liu, Min Liu, Shuai Jiang, Yi Tang, Yaonan Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments on three widely used semantic segmentation datasets: Cityscapes[34], Cam Vid [35], and BDD100K[36]... Extensive experiments demonstrate that DPS achieves state-of-the-art results within the same search space. The architectures discovered by DPS exhibit strong generalization, deliver superior performance, and maintain high efficiency. See Section 4 for experiments and results, including ablation studies and comparisons to other methods. |
| Researcher Affiliation | Academia | 1School of Artificial Intelligence and Robotics, Hunan University 2National Engineering Research Center of Robot Visual Perception and Control Technology 3Department of Data and Systems Engineering, The University of Hong Kong |
| Pseudocode | Yes | Algorithm 1: The optimization algorithm of TĪø. Algorithm 2: The optimization algorithm of qĻ. Algorithm 3: Dynamic Path Selection (DPS). Algorithm 4: Evolution Search. |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: See supplemental material. |
| Open Datasets | Yes | We conduct experiments on three widely used semantic segmentation datasets: Cityscapes[34], Cam Vid [35], and BDD100K[36]. |
| Dataset Splits | Yes | Cityscapes contains 2,975 training, 500 validation, and 1,525 test images... Cam Vid consists of 367 training, 101 validation, and 233 test images... BDD100K provides 7,000 training and 1,000 validation images... |
| Hardware Specification | Yes | All models are trained for 850 epochs with a total batch size of 12 across two RTX 3090 GPUs. Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The computer resources we used are listed in Appendix D. |
| Software Dependencies | No | The paper mentions using 'stochastic gradient descent (SGD) optimizer', 'polynomial decay policy', and 'OHEM loss' as part of the training strategy, but does not specify the versions of any underlying software frameworks (e.g., PyTorch, TensorFlow) or libraries used. |
| Experiment Setup | Yes | We employ the stochastic gradient descent (SGD) optimizer with an initial learning rate of 0.01, momentum of 0.9, and weight decay of 0.0005. The learning rate is adjusted using a polynomial decay policy with a power of 0.9. During training, we apply standard data augmentation techniques including random cropping, random scaling, and horizontal flipping. All models are trained for 850 epochs with a total batch size of 12 across two RTX 3090 GPUs. In the case of transfer learning, the learning rate is set to 0.007, and the models are trained for 500 epochs on Cam Vid [35] and 200 epochs on BDD100K [36], with all other settings remaining unchanged. Details of dynamic path selection are presented below. We summarize the key settings as follows: twin = 20, m = 10, k = 5, Ļµ1 = 2e-5, Ļµ2 = 4e-5. Details of the evolutionary search are presented below. We summarize the key settings as follows: M = 1, 500, G = 20, N = 50, pm = 0.2, ps = 0.1, k = 10. |