Segment Anything without Supervision

Authors: XuDong Wang, Jingfeng Yang, Trevor Darrell

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Evaluated across seven popular datasets, Un SAM achieves competitive results with the supervised counterpart SAM, and surpasses the previous state-of-the-art in unsupervised segmentation by 11% in terms of AR.
Researcher Affiliation Academia Xu Dong Wang Jingfeng Yang Trevor Darrell UC Berkeley
Pseudocode Yes Algorithm 1 Divide and Conquer
Open Source Code Yes code: https://github.com/frank-xwang/Un SAM
Open Datasets Yes MSCOCO [24], LVIS [15], SA-1B [21], ADE [48], Entity [29], Part Image Net [16] and PACO [30].
Dataset Splits No The paper mentions training on a percentage of SA-1B and evaluating on COCO Val2017 and SA-1B test set, but does not explicitly state validation dataset splits for its training process, apart from using standard evaluation sets which often serve as de-facto validation sets for benchmarking.
Hardware Specification Yes All model training in this paper was conducted using either 4 A100 GPUs or 8 RTX 3090 GPUs.
Software Dependencies No The paper mentions software components and models like DINO, Mask2Former, and Semantic-SAM, but does not provide specific version numbers for them or other dependencies.
Experiment Setup Yes In the divide stage, we set the confidence threshold τ=0.3; in the conquer stage, we choose threshold θmerge = [0.6, 0.5, 0.4, 0.3, 0.2, 0.1]. ... The default learning rate is 5e-5 with a batch size of 16 and a weight decay of 5e-2. We train the model for 8 epochs. ... The default learning rate is 1e-4 with a batch size of 8. The learning rate decreases by a factor of 10 at 90% and 95% of the training iterations. We train the model for 4 epochs.