PaintSeg: Painting Pixels for Training-free Segmentation
Authors: Xiang Li, Chung-Ching Lin, Yinpeng Chen, Zicheng Liu, Jinglu Wang, Rita Singh, Bhiksha Raj
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results demonstrate that Paint Seg outperforms existing approaches in coarse mask-prompt, box-prompt, and point-prompt segmentation tasks, providing a training-free solution suitable for unsupervised segmentation. |
| Researcher Affiliation | Collaboration | Xiang Li1, Chung-Ching Lin2, Yinpeng Chen2, Zicheng Liu2, Jinglu Wang2, Rita Singh1, Bhiksha Raj1,3 1CMU, 2Microsoft, 3MBZUAI |
| Pseudocode | No | The paper describes the AMCP process in detail (e.g., in section 4), but it does not include a formally labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Code: https://github.com/lxa9867/Paint Seg. |
| Open Datasets | Yes | For mask-prompt segmentation, we evaluate on DUTS-TE [63] and ECSSD [53]. ... For box-prompt segmentation, we evaluate on PASCAL VOC [19] val set and COCO [39] MVAL datasets. ... For point-prompt segmentation, we use three datasets including Grab Cut [50] which contains 50 images and corresponding segmentation masks that delineate a foreground object; Berkeley [42] which contains 96 images with 100 instances with more difficulty than Grab Cut and DAVIS [47]. |
| Dataset Splits | No | The paper mentions datasets used for evaluation but does not specify explicit train/validation/test splits by percentages or sample counts for reproducibility. |
| Hardware Specification | No | The paper does not explicitly mention the specific hardware (e.g., GPU model, CPU type) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'latent-diffusion pipeline [49]' and 'DINO [8] pretrained VIT-S/8 [17]', but it does not provide specific version numbers for these or other software components. |
| Experiment Setup | Yes | We set the diffusion iterations to 50. ... We set the number of cluster centers to 3 in the first three steps for point, box and scribble prompts otherwise 2. We set λpaint = 0.8, λcolor = 0.2 and λprompt = 0.2 if in I-step and λprompt = 0.2 if in O-step. We average N=5 painted images to obtain the updated mask for each step. The σx and σy are set to 1/10 of the width and height of the bounding box of the current stage mask respectively. + and are the neighbors 32 pixels outside and inside the object boundary. ... The kernel size is set to 5. ... The images are padded to 512x512 to fit the generative inpainting model. |