Differential Alignment for Domain Adaptive Object Detection

Authors: Xinyu He, Xinhui Li, Xiaojie Guo

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on widely-used DAOD datasets together with ablation studies are conducted to demonstrate the efficacy of our proposed method and reveal its superiority over other SOTA alternatives.
Researcher Affiliation Academia Xinyu He, Xinhui Li, Xiaojie Guo* College of Intelligence and Computing, Tianjin University, Tianjin, China xy EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes the methodology using textual explanations and mathematical equations, but it does not include a clearly labeled pseudocode block or algorithm section.
Open Source Code Yes Code https://github.com/Estrella-Xyu/Differential-Alignment-for-DAOD
Open Datasets Yes Cityscapes (Cordts et al. 2016) comprises 2,975 training images and 500 validation images, covering various urban environments and traffic conditions. [...] Foggy Cityscapes (Sakaridis et al. 2018) is a synthetic dataset rendered from Cityscapes with three levels of foggy density (0.005, 0.01, 0.02). [...] Sim10k (Johnson-Roberson et al. 2016) contains 10,000 images rendered from GTA engine. [...] BDD100k-daytime (Yu et al. 2020) is a subset of the larger BDD100k dataset, specially designed for daytime scenarios.
Dataset Splits Yes Cityscapes (Cordts et al. 2016) comprises 2,975 training images and 500 validation images, covering various urban environments and traffic conditions. [...] BDD100k-daytime (Yu et al. 2020) [...] It contains 36,728 training images and 5,258 validation images
Hardware Specification Yes Our implementation is based on the Py Torch framework and the model is trained on 4 NVIDIA RTX3090 GPUs with 24 GB of memory each.
Software Dependencies No The paper mentions "Py Torch framework" but does not specify a version number for PyTorch or any other software dependencies with their versions.
Experiment Setup Yes We optimize the network using the SGD optimizer with a momentum of 0.9. The initial learning rate is set to 0.01 and decreases in the final iteration. We use a batch size of 32, consisting of 16 labeled source images and 16 unlabeled target ones, and train the network for 25k iterations in total, including 10,000 iterations for burn-in and 15,000 iterations for teacher-student mutual learning. [...] In our experiments, setting α to 0.9996 works sufficiently well. [...] In our experiments, we set λ as 0.01 by default.