Hyperbolic Active Learning for Semantic Segmentation under Domain Shift
Authors: Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, Fabio Galasso
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experimental analysis based on two established synthetic-to-real benchmarks, i.e. GTAV Cityscapes and SYNTHIA Cityscapes. Additionally, we test HALO on Cityscape ACDC for domain adaptation under adverse weather conditions, and we benchmark both convolutional and attention-based backbones. HALO sets a new state-of-the-art in active learning for semantic segmentation under domain shift |
| Researcher Affiliation | Collaboration | *Equal contribution 1ITALAI S.R.L. 2Sapienza University of Rome 3Panasonic North America 4UC Berkeley. Correspondence to: Luca Franco <luca.franco@italailabs.com>, Paolo Mandica <paolo.mandica@uniroma1.it>. |
| Pseudocode | No | The paper does not contain any sections explicitly labeled 'Pseudocode' or 'Algorithm', nor are there any structured code-like blocks. |
| Open Source Code | Yes | Code available at https://github.com/ paolomandica/HALO. |
| Open Datasets | Yes | Datasets For pre-training, we utilize GTAV (Richter et al., 2016) and SYNTHIA (Ros et al., 2016) synthetic datasets... For ADA training and evaluation, we employ real-world target datasets, specifically Cityscapes (CS) train and val sets or ACDC train and test sets, both categorized into the same 19 classes. CS (Cordts et al., 2016) consists of 2,975 training samples and 500 validation samples. ACDC (Sakaridis et al., 2021) comprises 4,006 images captured under adverse conditions (fog, nighttime, rain, snow) to maximize the complexity and diversity of the scenes. |
| Dataset Splits | Yes | CS (Cordts et al., 2016) consists of 2,975 training samples and 500 validation samples. |
| Hardware Specification | Yes | For all experiments, the model is trained on 4 Tesla V100 GPUs using Py Torch (Paszke et al., 2019) and Py Torch Lightning with an effective batch-size of 8 samples (2 per GPU). |
| Software Dependencies | No | For all experiments, the model is trained on 4 Tesla V100 GPUs using Py Torch (Paszke et al., 2019) and Py Torch Lightning with an effective batch-size of 8 samples (2 per GPU). |
| Experiment Setup | Yes | Riemannian SGD optimizer with momentum of 0.9 and weight decay of 5e-4 is used for all the trainings. The base learning rates for the encoder and decode head are 1e-3 and 1e-2 respectively, and they are decayed with a polynomial schedule with power 0.5. The models are pre-trained for 15K iterations and adapted for an additional 15K on the target set. As per (Xie et al., 2022a), the source images are resized to 1280x720, while the target images are resized to 1280x640. |