haris018/SAM-3-BlindSpots

SAM3 Blind Spots Dataset

Overview

This dataset highlights failure cases and limitations (blind spots) observed while experimenting with the SAM3 segmentation model. The purpose of this dataset is to analyze scenarios where the model struggles to correctly segment objects when guided by text prompts.

The dataset includes different types of scenes such as:

• simple object detection
• complex multi-object scenes
• spatial reasoning
• action-based prompts
• camouflaged objects
• sports scenes

Each data point records:

• the image scenario
• the prompt given to the model
• the expected segmentation result
• the actual model output

These examples help illustrate where the model performs well and where it fails.

---

Model Tested

Model: facebook/sam3 Release Date: November 2025

SAM3 is a promptable image segmentation model designed to generate segmentation masks for objects in images based on prompts such as points, bounding boxes, or text descriptions.

---

How the Model Was Loaded

The model was tested using Kaggle with GPU support.

Example code used to load the model:

Testing Notebook attached

Dataset Structure

sam3-blindspots-dataset
│
├── images
│   ├── image1.jpg
│   ├── image2.jpg
│   ├── image3.jpg
│   └── ...
│
├── dataset.csv
├── testing_notebook.ipynb
│
└── README.md

---

Dataset Examples

Image #	Image Scenario	Prompt	Expected Output	Model Output
1	Dog sitting on a sofa	dog	dog segmented	Correctly segmented
2	Image with utensils on table	smallest size plate	smallest plate segmented	Incorrectly segmented
3	Bicycle partially hidden	bicycle	bicycle segmented	Correctly segmented
4	Complex table scene with many objects	ink pot	ink pot segmented	Not segmented
5	Table containing pen, book, cup, glasses	glasses	glasses segmented	Correctly segmented
6	Table containing pen, book, cup, glasses	pen	3 pens segmented	Correctly segmented
7	Two dogs (black and white)	white dog	white dog segmented	Correctly segmented
8	One man holding pen and paper while others discuss	man holding pen and paper	that person segmented	Two men segmented (one incorrect)
9	Speakers where one is less visible	speaker	speaker segmented	Correctly segmented
10	Partially visible chair	chair	chair segmented	Correctly segmented
11	One girl using laptop while others discuss	girl using laptop	only that girl segmented	Two girls segmented (one incorrect)
12	Camouflaged owl in environment	owl	owl segmented	Correctly segmented
13	Three people discussing; one girl holding a pen	girl holding a pen	girl segmented	Not segmented
14	Three people discussing; girl on left side	left side girl	that girl segmented	Not segmented
15	Camouflaged chameleon	chameleon	chameleon segmented	Not segmented
16	People celebrating a trophy	man looking at trophy	that person segmented	Not segmented
17	Cricket fielders and wicket keeper	cricket wicket keeper	the wicket keeper segmented	Not segmented
18	Players with one player bowing down	man bowing down	the person bowing down segmented	Not segmented
19	Players with one referee	football referee	referee segmented	Two men segmented (one incorrect)

---

Observed Blind Spots

1. Action-Based Understanding

The model struggles with prompts that describe actions or interactions between people and objects.

Examples:

• man holding pen and paper
• man looking at trophy
• girl holding a pen
• man bowing down

These require understanding actions, not just object presence.

---

2. Spatial Reasoning

The model fails when prompts involve relative spatial descriptions.

Example:

• left side girl

The model has difficulty identifying objects based on relative position within the scene.

---

3. Camouflaged Objects

Objects that blend with their surroundings are difficult for the model to segment.

Examples:

• chameleon
• camouflaged animals

This indicates a weakness in detecting low contrast or camouflaged objects.

---

4. Complex Scenes with Many Objects

In cluttered environments, the model sometimes fails to identify the correct object.

Example:

• ink pot among many objects on a table

---

5. Domain-Specific Objects (Sports Scenes)

The model struggles with sports-related roles or specific entities.

Examples:

• cricket wicket keeper
• football referee

These roles require understanding contextual roles rather than simple objects.

---

Recommended Datasets for Fine-Tuning

To address these limitations, the model could be fine-tuned on datasets containing:

Referring Expression Segmentation

Datasets that map natural language descriptions to object masks.

Examples:

• RefCOCO
• RefCOCO+
• RefCOCOg

---

Human Interaction Datasets

Datasets that capture human-object interactions and actions.

Examples:

• Visual Genome
• GQA

---

Camouflage Detection Datasets

Datasets designed for detecting camouflaged objects.

Examples:

• CAMO
• COD10K

---

Estimated Dataset Size for Improvement

A dataset of approximately 50,000 – 200,000 annotated samples would likely be needed to significantly improve performance.

Each sample should include:

• image
• natural language prompt
• segmentation mask

---

Purpose of This Dataset

This dataset is intended to:

• highlight limitations of segmentation models
• analyze failure cases of prompt-based segmentation
• support research on multimodal reasoning and segmentation models

---

License

This dataset is shared for research and educational purposes.