156 lines
4.5 KiB
Typst
156 lines
4.5 KiB
Typst
#import "@preview/postercise:0.2.0": *
|
|
#import themes.boxes: *
|
|
#import "@preview/fletcher:0.5.8" as fletcher: diagram, edge, node
|
|
|
|
#set page(width: 16in, height: 22in)
|
|
#set text(size: 16pt)
|
|
|
|
#show: theme.with(
|
|
primary-color: rgb(28, 55, 103), // Dark blue
|
|
background-color: white,
|
|
accent-color: rgb(243, 163, 30), // Yellow
|
|
titletext-color: white,
|
|
titletext-size: 1.8em,
|
|
)
|
|
|
|
#poster-header(
|
|
title: [Can SAM "Segment Anything"? #linebreak() ],
|
|
subtitle: [Evaluating Zero-Shot Performance on Crack Detection],
|
|
authors: [Hanwen Yu],
|
|
affiliation: [School of Advanced Technology, Supervisor: SiYue Yu
|
|
],
|
|
logo-2: image("./img/xjtlu-o.png", width: 15em),
|
|
)
|
|
|
|
|
|
// #image("examples.png", width: 100%)
|
|
|
|
|
|
#poster-content(col: 3)[
|
|
// Content goes here
|
|
#normal-box(color: none)[
|
|
== Introduction
|
|
he Segment Anything Model (SAM) has demonstrated remarkable
|
|
zero-shot segmentation capabilities on natural images. However, its zero-shot performance on domain-specific tasks remains underexplored.
|
|
// WHY CRACK SEGMENTATION?
|
|
// • Critical for infrastructure safety monitoring
|
|
// • Challenging characteristics:
|
|
// - Thin, elongated structures (often 1-5 pixels wide)
|
|
// - Low contrast against background
|
|
// - Complex branching topology
|
|
|
|
// RESEARCH QUESTION
|
|
|
|
*Can SAM2 achieve competitive crack segmentation
|
|
performance without domain-specific training?*
|
|
// CONTRIBUTIONS
|
|
// • First systematic evaluation of SAM2 zero-shot capability
|
|
// on crack segmentation
|
|
// • Comprehensive comparison of prompt strategies
|
|
// (bounding box vs. point-based prompts)
|
|
// • Analysis of failure modes and practical limitations
|
|
|
|
]
|
|
#normal-box(color: none)[
|
|
== Methodology
|
|
|
|
|
|
*Dataset*
|
|
|
|
- Crack500: 500 images with pixel-wise annotations
|
|
|
|
- Test set: 100 images for evaluation
|
|
|
|
*Prompt Strategies*
|
|
|
|
We evaluate four prompt generation approaches:
|
|
|
|
#table(
|
|
columns: 2,
|
|
|
|
[Prompt Type], [Description],
|
|
[Bounding Box], [Tight box around ground truth mask],
|
|
[1-Point Prompt], [Single point sampled from GT skeleton (morphological center)],
|
|
[3-Point Prompt], [Three uniformly distributed points along GT skeleton],
|
|
[5-Point Prompt], [Five uniformly distributed points along GT skeleton],
|
|
)
|
|
|
|
*Evaluation*
|
|
|
|
$
|
|
"IoU" = "TP" / ("TP" + "FP" + "FN")
|
|
$
|
|
|
|
$
|
|
"F1" = 2 * ("Precision" * "Recall") / ("Precision" + "Recall")
|
|
$
|
|
|
|
*Baselines*
|
|
|
|
Supervised models: UNet, DeepCrack, TransUNet,
|
|
CT-CrackSeg, VM-UNet, CrackSegMamba
|
|
|
|
|
|
#import fletcher.shapes: brace, diamond, hexagon, parallelogram, pill
|
|
|
|
#set text(size: 16pt)
|
|
#diagram(
|
|
node-fill: gradient.radial(white, blue, radius: 200%),
|
|
node-stroke: blue,
|
|
spacing: 25pt,
|
|
(
|
|
node((0, 0), [Crack Image], shape: rect),
|
|
node((0, 1), [SAM Image Encoder], shape: rect),
|
|
node((0, 2), [Prompt Generation #linebreak() BBox, 1/3/5 points], shape: rect),
|
|
node((1, 2), [SAM Mask Decoder], shape: rect),
|
|
node((1, 1), [Predircted Mask], shape: rect),
|
|
node((1, 0), [Metrics (IoU, F1)], shape: rect),
|
|
)
|
|
.intersperse(edge("-|>"))
|
|
.join(),
|
|
)
|
|
|
|
]
|
|
#normal-box(color: none)[
|
|
== Experiments and Results
|
|
#image("img/examples.png")
|
|
#image("img/metrics.png")
|
|
#image("img/sam_iou.png")
|
|
#image("img/sam_f1.png")
|
|
|
|
]
|
|
|
|
#normal-box(color: none)[
|
|
|
|
== Qualitative Analysis
|
|
#image("img/fail1.png")
|
|
#image("img/fail2.png")
|
|
|
|
]
|
|
|
|
#normal-box(color: none)[
|
|
== Key Findings and Discussion
|
|
// *Prompt Effectiveness*
|
|
|
|
Bounding box prompts yield the best performance among zero-shot methods. There is a 4.7x performance gap between bbox(39.6% IoU) and 1-point prompts(8.4% IoU).
|
|
|
|
|
|
SAM2 with bbox prompts (39.6% IoU) lags behind supervised models, even UNet in 2015. which highlights limitations of zero-shot approach without fine-tuning.
|
|
|
|
// *Single Point Prompt Limitations*
|
|
1-point prompts perform poorly (12.3% IoU), indicating insufficient guidance for complex crack structures. 5-point prompts approach bbox performance for highly irregular cracks, suggesting multiple points help capture shape.
|
|
]
|
|
|
|
#normal-box(color: none)[
|
|
|
|
== Conclusion and Future Work
|
|
|
|
SAM2 shows limited zero-shot capability for crack segmentation. Bounding box prompts significantly outperform point-based prompts. Performance still lags behind supervised methods, indicating need for domain adaptation.
|
|
]
|
|
#poster-footer[
|
|
// Content
|
|
Hanwen Yu | Email: Hanwen.Yu24\@student.xjtlu.edu.cn
|
|
]
|
|
]
|
|
|