Semantic & Instance Segmentation
Overcoming the Mask Proposal Recall Ceiling in Open-Vocabulary Segmentation
Generated automatically from the limitations stated in 9 papers (CVPR, ICCV, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current open-vocabulary and weakly supervised segmentation pipelines decouple class-agnostic mask proposal generation from segment classification, freezing an upstream proposal model like SAM, FreeSOLO, or Mask2Former. Because the downstream classifier operates solely on candidate masks, segmentation accuracy on novel classes and sub-object regions is strictly bounded by the proposal generator's recall, creating a ceiling that downstream improvements cannot fix. Failures to isolate novel or fine-grained instances at the proposal stage permanently propagate through the pipeline without any mechanism for recovery.
Why it matters
Enables segmentation frameworks to generalize beyond the fixed objecthood definitions and training class budgets of pre-trained proposal generators, segmenting fine-grained sub-parts and novel object categories previously missed.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Iterative Text-Guided Proposal Refinement: Construct an iterative mask adaptation module that takes text embeddings of candidate classes to modulate and deform frozen proposal masks, measuring mask average recall (AR) and panoptic quality (PQ) on novel classes in ADE20k/LVIS.
- 2
Multi-Granularity Candidate Ensembling: Design a lightweight merging and cross-scale pruning mechanism combining low-level unsupervised boundaries (e.g., self-supervised ViT patch features) with frozen proposal masks, evaluating recall gains across sub-object and rare entity benchmarks.
- 3
End-to-End Proposal-Classifier Co-adaptation: Train a lightweight feedback adapter that updates proposal boundaries based on downstream classification uncertainty, measuring the delta between fixed-proposal oracle upper bounds and adapted-proposal mIoU.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If foundation vision models (such as iterative promptable foundation segmenters) scale to near-100% zero-shot mask recall across all sub-object granularities natively, eliminating the proposal bottleneck via brute pretraining scale.
Evidence
Each paper's own statement of the limitation, verbatim.
- The Missing Point in Vision Transformers for Universal Image SegmentationCVPR 2026
Requires a frozen, already high-quality mask proposal generator; the method only improves classification, so failures in mask generation remain unaddressed
- SAM2Long: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory TreeICCV 2025
Dependency on the base model's ability to generate at least one viable candidate mask among its proposals
- Boosting Weakly Supervised Referring Image Segmentation via Progressive ComprehensionNeurIPS 2024
Requires pre-trained proposal generators (SAM or FreeSOLO) for the RaS and IaD losses and for final mask extraction, adding dependency and its own quality ceiling (oracle 69.0 mIoU bounds SAM-prompted accuracy)
- Open-Vocabulary Universal Image Segmentation with MaskCLIPICML 2023
Performance depends heavily on mask proposal quality: with GT masks PQ jumps from 15.1 to 35.8, so the class-agnostic proposal network is the bottleneck on unseen categories
- Open-Vocabulary Semantic Segmentation with Decoupled One-Pass NetworkICCV 2023
Performance is bounded by the class-agnostic mask proposal network, which the method does not itself improve; improvements rely on segment classification quality
- Open-Vocabulary Semantic Segmentation With Mask-Adapted CLIPCVPR 2023
Mask proposal generator is not truly class-agnostic — its notion of objecthood is bounded by COCO-Stuff's 171 class definitions, so fine-grained sub-object regions may not be segmented
- MasQCLIP for Open-Vocabulary Universal Image SegmentationICCV 2023
The mask proposal network is fixed once pre-trained, limiting capability for object types of arbitrary specification
- Cut and Learn for Unsupervised Object Detection and Instance SegmentationCVPR 2023
Masks are class-agnostic; the model has no notion of semantic classes and cannot be evaluated or used in class-aware detection setups
Show all 9 papers
- A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-Language ModelECCV 2022
Dependence on the mask proposal generator's ability to localize unseen classes correctly
Nearest existing work
- MasQCLIP for Open-Vocabulary Universal Image SegmentationICCV 2023
- A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-Language ModelECCV 2022
- FreeSeg: Unified, Universal and Open-Vocabulary Image SegmentationCVPR 2023
- Per-Pixel Classification is Not All You Need for Semantic SegmentationNeurIPS 2021
- Mask-Adapter: The Devil is in the Masks for Open-Vocabulary SegmentationCVPR 2025
- Open-Vocabulary Universal Image Segmentation with MaskCLIPICML 2023
- A Simple Framework for Open-Vocabulary Segmentation and DetectionICCV 2023
- Segment Anyword: Mask Prompt Inversion for Open-Set Grounded SegmentationICML 2025
- CLIP as RNN: Segment Countless Visual Concepts without Training EndeavorCVPR 2024
- Mask-Free OVIS: Open-Vocabulary Instance Segmentation Without Manual Mask AnnotationsCVPR 2023
- Towards Open-Vocabulary Semantic Segmentation Without Semantic LabelsNeurIPS 2024
- Open-Vocabulary Semantic Segmentation With Mask-Adapted CLIPCVPR 2023
- Scaling Open-Vocabulary Image Segmentation with Image-Level LabelsECCV 2022
- Masked-Attention Mask Transformer for Universal Image SegmentationCVPR 2022
- OpenMask3D: Open-Vocabulary 3D Instance SegmentationNeurIPS 2023
Related open problems
Semantic & Instance Segmentation
Benchmarking and Bridging Idealized Preconditions in Promptable Segmentation
Current promptable segmentation models systematically rely on idealized preconditions during evaluation and training, including ground-truth-derived prompts (points or tight boxes extracted directly from GT masks), dense instance annotations, and perfect spatial feature alignment. In real-world interactive deployment, prompts originate from noisy human clicks, coarse bounding boxes, or imperfect auxiliary detectors, leading to uncharacterized performance degradation. Furthermore, models optimized exclusively on point prompts fail to maintain performance across alternative prompt modalities or without high-resolution decoders. Consequently, reported benchmark gains reflect idealized prompting setups rather than actual operational capability.
Semantic & Instance Segmentation
Benchmarking and Developing Segmentation Without Dense Initial Annotation Preconditions
Current semantic and instance segmentation methods rely on strict supervision preconditions, such as dense pixel-wise ground truth or accurate initial bounding-box-derived masks. When applied to domains where only weak annotations (such as image-level tags or sparse points) are available from scratch, these methods cannot be initialized or trained. Consequently, segmentation models cannot be deployed in annotation-scarce domains without first undertaking costly manual pixel-level or box-level labeling campaigns.
Semantic & Instance Segmentation
Hierarchical Layout and Vector Segmentation Under Incomplete Multi-Level Supervision
Current instance and semantic segmentation methods for structured documents and vector graphics rely on full, multi-tier hierarchical annotations (simultaneously requiring word, line, and paragraph labels) and fail when vector elements are deeply nested, tangled, or overlapping. Because multi-level hierarchical annotations are significantly more expensive than standard single-level bounding boxes, deploying segmentation pipelines to diverse real-world documents remains cost-prohibitive. Furthermore, the inability to parse overlapping and nested primitives prevents reliable digitization of technical figures, CAD drawings, and complex vector assets.