Segment Any Change
Authors: Zhuo Zheng, Yanfei Zhong, Liangpei Zhang, Stefano Ermon
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments to confirm the effectiveness of Any Change for zero-shot change detection. Any Change sets a new record on the SECOND benchmark for unsupervised change detection, exceeding the previous SOTA by up to 4.4% F1 score, and achieving comparable accuracy with negligible manual annotations (1 pixel per image) for supervised change detection. |
| Researcher Affiliation | Collaboration | 1Stanford University 2Wuhan University |
| Pseudocode | Yes | The pseudo-code of Bitemporal Latent Matching is in Appendix B. |
| Open Source Code | Yes | Code is available at https://github.com/Z-Zheng/pytorch-change-models. |
| Open Datasets | Yes | Datasets: We use four commonly used change detection datasets to evaluate Any Change. The first three, i.e., LEVIR-CD (Chen & Shi, 2020), S2Looking (Shen et al., 2021), and x View2 (Gupta et al., 2019), are building-centric change detection datasets. SECOND (Yang et al., 2021) is a multi-class (up to 36 change types) urban change detection dataset with full annotation. |
| Dataset Splits | Yes | The optimal threshold is obtained by an adaptive threshold selection method, OTSU (Otsu, 1979). We also conduct a linear search for its optimal threshold on a small validation set, and the searched results have comparable performance with OTSU. |
| Hardware Specification | Yes | The forward computation was conducted on 8 NVIDIA A4000 GPUs. |
| Software Dependencies | No | The paper mentions general software like PyTorch and CUDA in the NeurIPS checklist, but does not specify their version numbers or other library versions required for replication in the main text or appendix sections on implementation details. |
| Experiment Setup | Yes | For object proposal generation, we adopt a point per side of 64, an NMS threshold of 0.7, a predicted Io U threshold of 0.5, and a stability score threshold is 0.8 for LEVIR-CD, S2Looking, SECOND, and 0.95 for x View2. [...] We only train a Lo RA (r = 32, alpha = 320, dropout = 0.1) for the SAM model on each dataset. [...] The training iterations are 200 epochs with a batch size of 16, Adam W optimizer with a weight decay of 0.01. The learning rate schedule is poly (γ = 0.9) decay with an initial learning rate of 6e-5 The training data augmentation adopts random rotation, flip, scale jitter, and cropping. The crop size is 512 for LEVIR-CD and S2Looking and 256 for x View2 and SECOND, respectively. |