Subject Taxonomy & Capture Strategy

The Decision Before the Shutter Matters 10× More

70% of 3DGS failures happen before capture, not during training. This chapter doesn't teach you how to press the shutter — it helps you see your subject clearly first. "Shooting a building" means completely different things for a three-story villa versus a traditional courtyard house.

Three Questions This Chapter Answers

1. What category does my subject belong to? What are the typical boundaries of this category?

2. Does my current equipment match this subject?

3. What's the most common failure for this subject type, and how do I avoid it in advance?

Five Subject Categories

All capturable subjects fall into 5 categories. Each has subtypes that determine specific strategy.

Category A · Scene

Definition: An environment where you need to walk through, look around, and experience parallax.

Category B · Object

Definition: A single independently observable object that can be captured by orbiting around it.

Category C · People & Living Subjects

Definition: Subjects that move, breathe, blink, or shift involuntarily.

Category D · Composite / Fragmented Space

Definition: Objects embedded in scenes, scenes containing featured objects, or non-contiguous spaces that need stitching.

Category E · Aerial / God's Eye View

Definition: Large-scale subjects requiring overhead coverage, captured by drone.

Key Drone Capture Parameters (2026 DJI Terra V5.2 Workflow):

DJI Terra V5.2 is the first mainstream tool in 2026 to integrate Gaussian Splatting fusion reconstruction into a drone workflow. It outputs both PLY (for downstream editing) and 3D Tiles (for Cesium web-based LOD streaming), supporting the full DJI Matrice 4E + Zenmuse P1 (45MP full-frame) pipeline.

Device × Subject Decision Matrix

Three Golden Rules (covers 80% of scenarios):

1. Phone for products and small objects

2. DSLR/mirrorless for scenes and building exteriors

3. Drone for building tops and large-area aerial views

Other combinations solve specific pain points — don't chase full equipment coverage from the start.

Complete Decision Matrix:

Difficulty: ▲ Easy / ▲▲ Medium / ▲▲▲ Hard / ✗ Not recommended

Radius & Trajectory: The Fundamental Difference Between Objects and Scenes

Objects: Spherical Orbit

Subject at center, camera follows a spherical trajectory, lowering elevation with each ring.

Ideal Coverage Parameters:

• Total: 50-150 photos

• Radius: 1.5-3× the object's longest dimension, held constant throughout

• Key principle: Move the camera, never the object. If the object moves, all previously captured photos are invalidated.

Scenes: Inward Trajectory

You're inside the scene, camera moves along interior paths, changing orientation at each stop.

Ideal Coverage Parameters:

Composite Spaces: Separate Then Combine

Split "exhibition hall + sculpture" into two capture sessions:

1. Capture the sculpture using object strategy (spherical orbit, 50-80 photos)

2. Capture the hall using scene strategy (inward trajectory, 100-200 photos)

3. Train the two datasets separately or merge for combined training

Never mix capture strategies. Mixed capture confuses SfM pose estimation — the algorithm cannot determine whether you're orbiting an object or walking through a scene.

Five Most Common Failures

Failure 1: Museum Display Case Glass

Glass registers as air during SfM (no feature points), but reflections become floating artifacts during training.

Solution:

• Capture two separate photo sets: before glass and behind glass, train separately

• Use a CPL polarizing filter to suppress reflections (rotate until reflections are darkest)

• Post-process in SuperSplat: manually remove Gaussians in the glass region

Failure 2: Sunset at the Beach — Continuously Changing Light

3DGS assumes constant scene illumination. During sunrise/sunset, color temperature shifts every 2-3 minutes. Passing clouds drop exposure by 2-3 stops instantly.

Solution:

• Complete all capture under constant lighting (overcast is ideal, or within 2 hours of noon)

• If sunset capture is unavoidable, finish all collection within 15 minutes

• Apply consistency correction in post using methods from 06-Color Grading

Failure 3: Forgot the Product's Bottom

30 photos of the front, zero of the bottom. Training produces a black hole underneath.

Solution:

• Every object capture must include ≥5 overhead/bottom-view shots

• Place the object on a transparent acrylic stand, shoot from below

• Even if you don't plan to display the bottom — capture it anyway. The void will "infect" adjacent areas

Failure 4: Forgot the Building's Top

200 photos from ground level, but no angles on the roof or under the eaves. The entire upper portion collapses into blur.

Solution:

• Use a drone for top-down and 45° oblique angles (DJI Mini 4 Pro is sufficient)

• Without a drone, use 70-200mm telephoto from a distance to capture upper details

• Shoot from upper-floor windows of an adjacent building (if accessible)

Failure 5: 360° Camera on a Small Object

360° cameras (e.g., Insta360 X4) have a minimum focus distance of ~60cm. Objects under 30cm will be severely out of focus, appearing as blurry textures.

Solution:

• 360° cameras suit large spaces (≥3m × 3m), not objects

• For small objects, use a phone or DSLR + 50mm/macro lens

• KIRI Engine's object scanning mode is specifically optimized for guiding capture of objects <50cm

Example: Minimum Dataset Structure for Objects

my-object-2026-05/ ├── raw/
├── DSC0001.ARW │
├── DSC0001.JPG │
└── ... (52 files) ├── selected/
# Curated 50 photos (manually filtered) │
├── 001.jpg
# Naming: three-digit sequential │
├── 002.jpg │
└── ... (50 files) ├── masks/
# Optional: background removal alpha masks │
├── 001_mask.png │
└── ... └── meta.yaml
# Capture metadata

Recommended meta.yaml fields:

subject: "Song Dynasty celadon bowl" date: 2026-05-07 device: "Sony A7M4 + FE 35mm f/1.8" lighting: "single softbox, 5600K daylight balance" radius_meters: 0.6 photo_count: 52 orbit_rings: 3 angles_per_ring: [18, 20, 14] notes: "6 overhead shots of bottom; front color temp 200K cool, batch corrected in Lightroom"

Next Steps

• Ready to start capturing → Enter 02-Scouting & Capture Planning

• Want camera parameters first → Enter 03-Camera Parameters & Field Operations

• See real results → Gallery · Creator Picks

• Want to skip SfM with cloud processing → Enter 08-Training