Order-aware Interactive Segmentation
Authors: Bin Wang, Anwesa Choudhuri, Meng Zheng, Zhongpai Gao, Benjamin Planche, Andong Deng, Qin Liu, Terrence Chen, Ulas Bagci, Ziyan Wu
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that OIS achieves state-of-the-art performance, improving m Io U after one click by 7.61 on the HQSeg44K dataset and 1.32 on the DAVIS dataset as compared to the previous state-of-the-art Seg Next, while also doubling inference speed compared to current leading methods. We evaluate our method and comparison methods on two widely used benchmarks for interactive segmentation: HQSeg44K (Ke et al., 2024) and DAVIS (Perazzi et al., 2016). Table 1: Performance comparison on the HQSeg44K benchmark. Table 4: Ablation experiments on DAVIS. |
| Researcher Affiliation | Collaboration | 1Northwestern University, Chicago, IL, USA 2United Imaging Intelligence, Boston, MA, USA 3University of Central Florida, Orlando, FL, USA 4University of North Carolina at Chapel Hill, Chapel Hill, NC, USA |
| Pseudocode | No | The paper describes the methodology in text and uses diagrams, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | To ensure the reproducibility of our work, we extensively describe the implementation details, such as the pre-trained model, training data, and other important training hyperparameters in Sec 3, Sec 4.1 and Sec A.1. We will release all source code pending release approval. |
| Open Datasets | Yes | We evaluate our method and comparison methods on two widely used benchmarks for interactive segmentation: HQSeg44K (Ke et al., 2024) and DAVIS (Perazzi et al., 2016). |
| Dataset Splits | No | We evaluate our method and comparison methods on two widely used benchmarks for interactive segmentation: HQSeg44K (Ke et al., 2024) and DAVIS (Perazzi et al., 2016). More details about these two datasets can be found in Appendix A.2. To be consistent with previous works, we use a subset of 345 frames to conduct the evaluation. |
| Hardware Specification | Yes | We use Adam optimizer to train our model on HQSeg44K dataset for 15 epochs on two A100 GPUs. |
| Software Dependencies | No | The paper mentions using an "Adam optimizer" and a "frozen Vi T-Base encoder from Depth Anything V2" but does not provide specific version numbers for these or other key software components (e.g., PyTorch, TensorFlow, specific Python libraries). |
| Experiment Setup | Yes | The input image size is set to 1024. We set the maximum number of clicks N as 48 during training, with the first 24 for positive clicks and the remaining for negative. We use three same blocks of object-level and order-level understanding modules (Sec. 3.2 and Sec. 3.3) to conduct prompt fusion. In these modules, the Feed-Forward Network (FFN) is implemented as a 2-layer Multi-Layer Perceptron (MLP). Each attention module is followed by a Layer Norm, which normalizes the sparse embeddings. We use Adam optimizer to train our model on HQSeg44K dataset for 15 epochs on two A100 GPUs. |