PaperlayerView exact revision on arXiv

Source-linked AI summary

Segment Anything

Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, Ross Girshick

arXiv:2304.02643v1cs.CVcs.AIcs.LG

TL;DR

The project asks which task, model, and data can support zero-shot generalization in image segmentation. It introduces promptable segmentation, SAM, and SA-1B, achieving consistently strong zero-shot results across downstream tasks.

  • Problem

    The paper asks what task, model architecture, and data can enable zero-shot generalization for image segmentation.

  • Method

    The project introduces promptable segmentation, the SAM model, and the SA-1B dataset built through an efficient data-collection loop.

  • Results

    SAM produces high-quality masks and consistently strong zero-shot results across downstream tasks including edge detection, object proposals, instance segmentation, and text-to-mask prediction.

  • Takeaways & Limitations

    Prompting and composition make SAM usable with other components for extensible applications beyond systems trained for fixed tasks.

  • Takeaways & Limitations

    SAM can miss fine structures, hallucinate small disconnected components, and produce less crisp boundaries than more computationally intensive methods.

Abstract

from arXiv · show

We introduce the Segment Anything (SA) project: a new task, model, and dataset for image segmentation. Using our efficient model in a data collection loop, we built the largest segmentation dataset to date (by far), with over 1 billion masks on 11M licensed and privacy respecting images. The model is designed and trained to be promptable, so it can transfer zero-shot to new image distributions and tasks. We evaluate its capabilities on numerous tasks and find that its zero-shot performance is impressive -- often competitive with or even superior to prior fully supervised results. We are releasing the Segment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and 11M images at https://segment-anything.com to foster research into foundation models for computer vision.

1. Introduction

The work proposes a foundation model for image segmentation built around a promptable task, architecture, and large-scale data engine. It introduces SAM and SA-1B, evaluates zero-shot transfer and fairness, and releases both for research use.

  • Motivation: The project seeks a promptable image-segmentation foundation model that pre-trains broadly and generalizes to downstream problems on new data distributions.Its design is organized around task, model, and data.
  • Task: The proposed task returns a valid segmentation mask for any segmentation prompt, including spatial or text information specifying what to segment.The task is intended to support zero-shot generalization and diverse downstream applications.
  • Model: SAM combines an image encoder, prompt encoder, and mask decoder to support flexible prompts, amortized real-time mask computation, and ambiguity awareness.These architectural constraints arise from interactive real-world use of the promptable segmentation task.
  • Data engine: The data engine co-develops SAM with model-in-the-loop annotation because strong generalization requires large and diverse masks unavailable in existing segmentation datasets.This strategy addresses the scarcity of naturally abundant segmentation masks.
  • Dataset: SA-1B contains more than 1B masks from 11M licensed and privacy-preserving images, providing 400× more masks than any existing segmentation dataset.The dataset was collected fully automatically in the final data-engine stage, with high quality and diversity verified through human ratings and experiments.
  • Experiments and release: SAM is evaluated across 23 segmentation datasets and downstream zero-shot tasks, while the work studies fairness and releases SAM under Apache 2.0 with SA-1B for research.Reported downstream tasks include edge detection, object proposal generation, and instance segmentation.

2. Segment Anything Task

Segment Anything defines promptable segmentation as returning a valid mask for any prompt, including ambiguous prompts that may refer to multiple objects. This task supports simulated-prompt pre-training and zero-shot transfer, allowing one model to adapt to diverse segmentation tasks through prompt engineering and composition.

  • Task: Promptable segmentation returns a valid mask from foreground/background points, rough boxes or masks, free-form text, or other information indicating what to segment.A valid mask remains required when an ambiguous prompt could refer to multiple objects.
  • Pre-training: Pre-training simulates sequences of prompts for each training sample and compares predicted masks with ground truth, adapting an interactive-segmentation method.Unlike interactive segmentation, the aim is to predict a valid mask for any prompt rather than only after enough user input.
  • Zero-shot transfer: Prompt engineering enables zero-shot transfer by casting downstream segmentation tasks as prompts, such as using cat detector boxes for cat instance segmentation.The task is intended to support a wide array of practical segmentation tasks through appropriate prompts.
  • Related tasks: The model targets task generalization across many existing and new segmentation tasks, unlike multi-task systems trained and tested on a fixed set of the same tasks.The supplied passage contrasts promptable adaptation with joint semantic, instance, and panoptic segmentation systems.
  • Discussion: Prompting and composition make a single model extensible to applications and tasks unknown at design time, motivating composable systems analogous to other foundation-model uses.The passage illustrates this analogy with CLIP serving as the text-image alignment component of DALL·E.

3. Segment Anything Model

SAM is a promptable segmentation model built from an image encoder, flexible prompt encoder, and fast mask decoder. Its design supports high-resolution inputs, ambiguous prompts, and efficient interactive use.

  • Architecture: SAM comprises an image encoder, flexible prompt encoder, and fast mask decoder designed for amortized real-time performance.The model builds on Transformer vision models.
  • Image encoder: The image encoder uses an MAE-pre-trained ViT minimally adapted for high-resolution inputs and runs once before prompting.Its embedding can be reused across prompts for the same image.
  • Prompt encoder: The prompt encoder represents sparse points, boxes, and text alongside dense masks, combining positional encodings, learned embeddings, CLIP features, and convolutions.Sparse prompt representations use positional encodings and prompt-type embeddings, while dense masks are embedded with convolutions.
  • Resolving ambiguity: SAM predicts 3 output masks for ambiguous prompts and ranks them using confidence scores estimating IoU.Multiple outputs address common nested-mask cases, and training backpropagates only through the minimum loss over masks.
  • Efficiency: ∼50ms is the prompt encoder and mask decoder runtime in a web browser on CPU given a precomputed image embedding.This enables seamless, real-time interactive prompting.
  • Losses and training: SAM is trained with focal and dice losses using geometric prompts sampled randomly in 11 rounds per mask.The simulated interactive setup allows the model to integrate with the data engine.

4. Segment Anything Data Engine

The SA-1B dataset was collected through a three-stage data engine progressing from model-assisted manual annotation to semi-automatic and fully automatic mask generation. As the model improved through retraining and architectural scaling, annotation became substantially faster, while later stages increased mask diversity and enabled automatic handling of ambiguity.

  • Data engine stages: The data engine collected the 1.1B-mask SA-1B dataset through model-assisted manual, semi-automatic, and fully automatic stages.The stages progressively reduced annotator input, ending with model-generated masks without annotator input.
  • Assisted-manual stage: Annotators used SAM’s browser-based interactive tool, clicking foreground/background points and refining masks with pixel-precise brush and eraser tools.Precomputed image embeddings enabled real-time model-assisted annotation directly in the browser.
  • Assisted-manual stage: 14 seconds was the average annotation time per mask after six retrainings, down from 34 seconds as the encoder scaled from ViT-B to ViT-H.The model began with public segmentation datasets and was later retrained using newly annotated masks.
  • Semi-automatic stage: The semi-automatic stage targeted less prominent objects by prefilling confident masks and asking annotators to label additional unannotated objects.A bounding box detector trained on first-stage masks identified confident masks using a generic “object” category.
  • Fully automatic stage: The fully automatic stage became feasible after model improvement and the development of an ambiguity-aware model that predicted valid masks in ambiguous cases.The model was prompted with a 32×32 regular grid of points.

5. Segment Anything Dataset

SA-1B is a large-scale segmentation dataset containing 11M licensed, privacy-protecting images and 1.1B masks, collected largely automatically with a data engine. Its automatic masks are high quality, while its image and mask coverage substantially exceeds existing segmentation datasets.

  • Dataset overview: SA-1B contains 11M diverse, high-resolution, licensed, privacy-protecting images and 1.1B high-quality segmentation masks.The dataset is released to support future foundation-model research in computer vision under a favorable license for certain research uses, with researcher protections.
  • Mask generation: 99.1% of SA-1B’s 1.1B masks were generated fully automatically by the data engine.Because most masks were automatic, the paper evaluates their quality against professional annotations and finds them effective for training models.
  • Mask quality: 94% of automatically predicted masks had greater than 90% IoU with professionally corrected masks, while 97% exceeded 75% IoU.Professional annotators corrected masks from a random sample of 500 images containing approximately 50k masks using pixel-precise editing tools.
  • Dataset scale: SA-1B has 11× more images and 400× more masks than Open Images, the largest existing segmentation dataset.The dataset’s scale is compared with prominent segmentation datasets using image and mask counts.
  • Mask properties: SA-1B covers image corners more extensively than LVIS v1 and ADE20K, whereas COCO and Open Images V5 show a stronger center bias.All compared datasets exhibit common photographer biases in the spatial distribution of object centers.

6. Segment Anything RAI Analysis

The RAI analysis examines geographic and income representation in SA-1B and fairness across protected attributes in SAM. It also provides dataset, annotation, and model cards to document these considerations.

  • RAI analysis: The RAI analysis investigates geographic and income distributions in SA-1B and fairness across protected attributes when using SA-1B and SAM.The fairness analysis covers perceived gender presentation, age group, and skin tone.
  • RAI analysis: The project provides dataset, data annotation, and model cards as additional documentation for its responsible-AI analysis.These materials are provided in §F.
  • Geographic and income representation: SA-1B has higher representation in Europe, Asia & Oceania, and middle income countries, while Africa, Latin America & Caribbean, and low income countries are underrepresented across datasets.The comparison includes SA-1B, COCO, and Open Images.
  • Fairness in segmenting people: All confidence intervals overlap across the evaluated groupings except between older and middle age groups.Table 2 reports 95% confidence intervals for segmenting people across perceived gender presentation, age group, and skin tone.
  • Fairness in segmenting people: SAM’s fairness evaluation measures performance discrepancies across perceived gender presentation, age group, and skin tone using MIAP and a proprietary dataset.It uses simulated interactive segmentation with random sampling of 1 and 3 points.

7. Zero-Shot Transfer Experiments

SAM is evaluated on five zero-shot transfer tasks and datasets unseen during training, including novel image distributions. Across promptable segmentation and higher-level tasks, it produces strong qualitative and quantitative results, though performance reflects task and annotation biases.

  • Experimental scope: SAM is tested on five tasks, four substantially different from promptable segmentation, using unseen datasets that may include underwater and egocentric images.The experiments assess transfer beyond the training task and distributions represented in SA-1B.
  • Single-point segmentation: SAM produces valid masks from ambiguous single foreground-point prompts, evaluated with both mIoU and human quality ratings.A single point can refer to multiple objects, making automatic metrics unreliable when ground-truth masks do not enumerate all possibilities.
  • Single-point segmentation: 16 of 23 datasets show higher mIoU for SAM than RITM, with gains reaching ∼47 IoU.An oracle selecting the most relevant of SAM’s three masks reveals how ambiguity affects automatic evaluation.
  • Single-point segmentation: SAM’s human-rated masks consistently surpass RITM, with mean ratings between 7 and 9, including datasets where its automatic metrics are worse.The ambiguity-unaware single-mask variant scores lower than SAM but still exceeds RITM.
  • Edge detection: Although untrained for edge detection, SAM produces reasonable edge maps with high R50 at the cost of precision, while trailing specialized state-of-the-art methods.SAM also predicts sensible edges absent from BSDS500 annotations.
  • Object proposal evaluation: SAM outperforms ViTDet-H on medium, large, rare, and common objects in proposal evaluation, but underperforms on small and frequent objects.ViTDet-H performs best overall when its detections are used as object proposals, benefiting from LVIS-specific annotation biases.
  • Instance segmentation: SAM’s mask AP is reasonably close to ViTDet on COCO and LVIS, while human inspection finds often-crisper SAM boundaries despite dataset-specific annotation biases.The comparison attributes some performance gaps to learned COCO and LVIS mask conventions that zero-shot SAM cannot exploit.
  • Text-prompted segmentation: With text-aware training that adds no new text annotations, SAM segments objects from simple prompts and phrases, and an extra point often fixes failures.Examples include “a wheel” and “beaver tooth grille.”

8. Discussion · Appendix

The discussion positions SAM as a broadly adaptable, composable foundation-model-style system for promptable segmentation while acknowledging limitations in fine-detail accuracy and interactive performance. The appendix documents implementation, dataset, annotation, model-card, RAI, and human-study details.

  • 8. Discussion: SAM aligns with the foundation-model paradigm because it is intended to be trained on broad data at scale and adapted across downstream tasks.The paper notes that foundation models are trained on broad data at scale and adaptable to a wide range of downstream tasks.
  • 8. Discussion: SAM seeks straightforward composition with other systems by predicting valid masks across a wide range of segmentation prompts.This broad prompt coverage is intended to provide a reliable interface between SAM and other components.
  • 8. Discussion: SAM can miss fine structures, hallucinate small disconnected components, and produce less crisp boundaries than more computationally intensive zoom-in methods.The discussion states that SAM is not perfect despite performing well in general.
  • 8. Discussion: Dedicated interactive segmentation methods are expected to outperform SAM when many points are provided, because SAM prioritizes generality and breadth over high-IoU interactive segmentation.This limitation reflects SAM’s design objective rather than a focus on maximizing interactive segmentation accuracy.
  • 8. Discussion: The project contributes a new task, promptable segmentation, a model, SAM, and a dataset, SA-1B, to advance image segmentation toward foundation models.The conclusion describes these as the principal contributions making this transition possible.
  • 8. Discussion: The conclusion leaves SAM’s foundation-model status to community use while expecting its perspective, release of over 1B masks, and promptable model to advance research.The paper explicitly presents this status as unresolved.
  • Appendix: The appendix covers SAM and task details, automatic mask generation, RAI, experiment implementation, human-study design, dataset and model cards, and annotation guidelines.These topics are organized across Sections A through G.

A. Segment Anything Model and Task Details · B. Automatic Mask Generation Details

SAM combines a high-resolution ViT image encoder, prompt encoders, and a lightweight mask decoder to produce prompt-conditioned masks, including multiple outputs for ambiguous prompts. Its automatic generation pipeline uses dense point-grid cropping, confidence and stability filtering, non-maximum suppression, and postprocessing to create SA-1B masks.

  • A. Segment Anything Model and Task Details: SAM uses an MAE-pretrained ViT-H/16 image encoder with windowed and global attention, producing a 16× downscaled image embedding.Inputs are rescaled and padded to 1024×1024, yielding a 64×64 embedding that is projected to 256 channels.
  • A. Segment Anything Model and Task Details: Sparse point and box prompts receive 256-dimensional embeddings, while dense mask prompts are downscaled and projected before being added to the image embedding.A learned embedding represents the absence of a mask prompt.
  • A. Segment Anything Model and Task Details: The lightweight decoder alternates token self-attention, token-to-image cross-attention, token MLP updates, and image-to-token cross-attention before dynamically predicting masks.Positional encodings and original prompt tokens are repeatedly re-added to preserve geometric location and prompt type.
  • A. Segment Anything Model and Task Details: SAM predicts three masks by default for ambiguous single prompts and returns only a dedicated fourth-mask prediction when multiple prompts are provided.The design addresses multiple valid masks while avoiding degenerate losses and preserving a regular gradient signal for unambiguous masks.
  • A. Segment Anything Model and Task Details: Training combines focal and dice losses in a 20:1 ratio and simulates interactive segmentation with an initial point or noisy box followed by error-driven prompts.The model also receives the previous iteration’s unthresholded mask logits as an additional prompt.
  • B. Automatic Mask Generation Details: Automatic generation samples masks from a 32×32 full-image point grid and 20 zoomed-in crops, then applies box-based non-maximum suppression.The crops use 2×2 and 4×4 partially overlapping windows with 16×16 and 8×8 point grids.
  • B. Automatic Mask Generation Details: The generation pipeline retains masks with predicted IoU at least 88.0 and stability IoU at least 95.0, filtering for confidence and threshold stability.The retained binary mask is obtained by thresholding logits at 0 after comparing masks thresholded at -1 and +1.
  • B. Automatic Mask Generation Details: Postprocessing removes connected components and holes smaller than 100 pixels, addressing estimated 4% rates for each spurious error type.Entire masks are removed when their largest component is below the component threshold.

C. RAI Additional Details

The additional analyses infer approximate geographic and income attributes from image captions and countries, and examine SAM’s segmentation performance across demographic attributes using external annotations. Clothing segmentation is evaluated on 6.5k Open Images masks linked to MIAP person boxes.

  • Geographic and income information: Caption-based named entity recognition assigns each SA-1B image an approximate country, province, and city location.An Elmo-based model extracts location entities, which are matched hierarchically to countries, provinces, and cities.
  • Geographic and income information: Flickr metadata provides approximate locations for 19,562 COCO training images and 493,517 Open Images training images with masks.These represent 24% of COCO and 18% of Open Images training data, respectively.
  • Geographic and income information: Image countries are mapped to World Bank income levels, with upper-middle and lower-middle categories collapsed into one middle level.Income information is derived from each image’s inferred country.
  • Fairness in segmenting people: SAM’s fairness on people is assessed across perceived gender presentation and age group using MIAP annotations and Open Images person-category masks.MIAP supplies box annotations, which are used to select corresponding Open Images masks for ground-truth analysis.
  • Fairness in segmenting clothing: Clothing fairness is evaluated on 6.5k Open Images ground-truth masks under the clothing superclass within MIAP person boxes, comparing perceived gender presentation and age group.Table 6 reports clothing-segmentation performance across these attributes; gender intervals are disjoint, while age-group confidence intervals overlap.

D. Experiment Implementation Details … D.6. Probing the Latent Space of SAM

The experiments evaluate SAM’s zero-shot transfer across diverse segmentation, edge detection, object proposal, instance segmentation, and text-to-mask settings, while also probing whether its latent representations capture shape and semantic similarity. The implementation uses task-specific prompting and evaluation protocols, and the latent-space results are preliminary but suggest downstream utility.

  • D.1. Zero-Shot Single Point Valid Mask Evaluation: SAM’s point-prompt evaluation measures IoU after 1, 2, 3, 5, or 9 iterative prompts, averaging per-mask IoU within datasets and then across all 23 datasets.Points are sampled either deterministically from object boundaries and error regions or randomly when specified; oracle evaluation selects the best of three predictions at one point.
  • D.1. Zero-Shot Single Point Valid Mask Evaluation: A 23-dataset benchmark spans diverse image domains and subsamples larger datasets to approximately 10k masks for efficient zero-shot evaluation.The benchmark includes egocentric, microscopy, X-ray, underwater, aerial, simulation, driving, and painting images.
  • D.2. Zero-Shot Edge Detection: On BSDS500, zero-shot edge detection uses a 16×16 foreground-point grid, mask deduplication, and Sobel filtering, producing qualitatively sensible edges despite no edge-detection training.Results are reported on the 200-image test subset using ODS, OIS, AP, and R50; predictions can align well with human annotations but often contain more edges than ground truth.
  • D.3. Zero-Shot Object Proposals: For LVIS zero-shot object proposals, SAM generates approximately 900 masks per image from a 64×64 point grid with a 0.9 NMS threshold and is evaluated using AR@1000.The comparison uses cascade ViTDet-H, while an ablation tests whether SAM’s three masks per point improve recall over single-output SAM.
  • D.4. Zero-Shot Instance Segmentation: For zero-shot instance segmentation, SAM refines masks generated from fully supervised ViTDet-H box prompts on COCO and LVIS validation splits, with cleaner boundaries reported qualitatively.The most confident predicted mask is fed back with the box prompt for an additional refinement iteration.
  • D.5. Zero-Shot Text-to-Mask: Zero-shot text-to-mask training uses normalized ViT-L/14@336px CLIP embeddings and 120k iterations, while inference supplies unmodified CLIP text embeddings as SAM prompts.Training masks come from the first two data-engine stages, masks smaller than 100² pixels are discarded, and prompts are generated from randomized expanded crops with optional masking.
  • D.6. Probing the Latent Space of SAM: A qualitative latent-space probe computes mask embeddings from cropped, horizontally flipped images and binary-mask-weighted image features, finding imperfect nearest-neighbor shape and semantic similarity.The preliminary results suggest possible uses in data labeling, dataset-content understanding, and downstream-task features.

E. Human Study Experimental Design · F. Dataset, Annotation, and Model Cards

The human study evaluates predicted mask quality through professional ratings of object validity, boundary cleanliness, and input correspondence, addressing limitations of IoU-based assessment. The paper also documents the SA-1B dataset, annotation process, and model through dedicated cards.

  • E. Human Study Experimental Design: The study compared RITM, single-output SAM, and SAM on single-point inputs, and cascade ViTDet-H with SAM for instance segmentation.The comparisons tested whether SAM improved visual quality and whether disambiguating multiple masks helped with ambiguous point prompts.
  • E. Human Study Experimental Design: Seven datasets spanning scene-level, ego-centric, drawn, overhead, underwater, and synthetic imagery were selected for single-point review.The selected datasets were LVIS v0.5, VISOR, DRAM, IBD, NDD20, OVIS, and iShape.
  • E. Human Study Experimental Design: Professional annotators rated each mask from 1–10 for object validity, boundary cleanliness, and correspondence to the model input.Each job included the image, one model’s predicted mask, and either a single point or single box.
  • E. Human Study Experimental Design: 1,000 masks per dataset were randomly sampled for single-point experiments, while 1,000 masks came from LVIS v1 validation for instance segmentation.Single-point inputs used centermost mask points; instance-segmentation inputs used LVIS ground-truth boxes.
  • E. Human Study Experimental Design: 4,000 rating jobs per dataset covered RITM, single-output SAM, SAM, and ground truth, while instance segmentation used 3,000 jobs for ViTDet, SAM, and ground truth.Jobs were randomly ordered, and 30 annotators drew jobs; initial five-rater testing produced an average score standard deviation of 0.83.
  • E. Human Study Experimental Design: All statistical tests comparing SAM with baseline models and single-output SAM were strongly significant, using paired t-tests and 99% confidence intervals from paired bootstrap tests on 10k samples.The reported confidence intervals excluded zero.
  • E. Human Study Experimental Design: The instance-segmentation comparison included 794 COCO ground-truth ratings and subsampled LVIS results to the same 794 inputs for fair comparison.Table 8 used the full 1000 ratings.
  • F. Dataset, Annotation, and Model Cards: The paper provides a Dataset Card for SA-1B, a Data Annotation Card for the first two data-engine stages, and a Model Card in Table 9.The cards follow the referenced Dataset Card, CrowdWork-Sheets, and Model Card frameworks.

F.1. Dataset Card for SA-1B

SA-1B is a large-scale segmentation dataset of 11M photos and approximately 1.1B masks, created to support computer-vision research. It protects privacy by blurring faces and license plates, while retaining limitations from automatically generated masks, withheld captions, and potentially sensitive imagery.

  • Dataset purpose and composition: SA-1B contains 11M photos and approximately 1.1B segmentation masks, making it the largest segmentation dataset to date.The photos cover varied subject matter, including locations, objects, and scenes.
  • Privacy and safety: Privacy protection includes blurring faces and license plates in all images, with no intended direct identification of individuals.Users can report anonymization issues to the dataset maintainers.
  • Data acquisition: The images were licensed from a third-party photo provider, while approximately 2k randomly selected images were withheld for testing.The remaining licensed images were included in the dataset.
  • Annotations: Each image is annotated with masks but has no categories or text labels, and the average image contains approximately 100 masks.The masks are generated automatically by a segmentation model.
  • Limitations and risks: Known limitations include erroneous or inconsistent generated masks, unavailable captions, and scenes that may reveal sensitive information such as protests or religious gatherings.The dataset also includes some closely timed photos of the same subject despite having no duplicate images.

F.2. Data Annotation Card · G. Annotation Guidelines

The SA-1B annotation process addressed subjective segmentation through experienced annotators, visual guidelines, iterative feedback, and quality checks. Its mask-review guidelines specified concrete consistency and boundary-error criteria, while the dataset was produced by 130 Kenya-based annotators without sociodemographic selection.

  • F.2. Data Annotation Card: The research team completed 30 annotation tasks to identify challenges, resolve complex cases, and refine visual annotation guidelines.
  • F.2. Data Annotation Card: Annotators were instructed to segment every identifiable object, using corrective clicks or bounding boxes and pixel-precise refinement tools.
  • F.2. Data Annotation Card: Annotators were selected for prior experience with other vision annotation tasks, and no sociodemographic characteristics were used in selection.
  • F.2. Data Annotation Card: 130 annotators, all based in Kenya, created the annotations; the authors reported no meaningful sociodemographic impact on the data.
  • F.2. Data Annotation Card: Weekly reviews communicated common mistakes and corrections, while the annotation QA team provided daily improvement feedback.
  • F.2. Data Annotation Card: Mask quality was reviewed through five complementary image views and a 1–10 rating interface for each mask-image pair.
  • G. Annotation Guidelines: The complete human mask-quality review guidelines were provided in Figures 19 and 20, with edited images and blurred faces for release.
  • G. Annotation Guidelines: Review examples marked combining unrelated objects, inconsistent handling of obscuring foreground objects, and disagreement with a provided point as mask errors.
Loading 2304.02643v1…