Perf: vectorize _get_pairwise_iou with a joint bincount

[aymuos15]

Description

Replaces the nested per-pair loop in _get_pairwise_iou (which held T + P full-volume masks live
and did O(pairs * volume) work) with a single joint bincount over gt * (num_pred + 1) + pred,
computing all pairwise intersections in one O(volume) pass that is much faster and, for realistic
instance counts, lower in peak memory.

Benchmark

Synthetic nuclei-style masks (random blobs), CPU, OMP_NUM_THREADS=8, median of repeated runs.

2D

Input (H×W, #true=#pred)	Speedup	Peak-mem saved
128×128, 40	~11×	83%
256×256, 80	~19.5×	95%
384×384, 120	~19.9×	95%
512×512, 50	~6×	96%
512×512, 150	~18.3×	97%
512×512, 300	~44.8×	98%
640×640, 180	~21.8×	97%
768×768, 200	~23.1×	96%
1024×1024, 150	~16.1×	95%

3D

Input (D×H×W, #true=#pred)	Speedup	Peak-mem saved
64×64×64, 30	~4.4×	92%
96×96×96, 50	~4.6×	86%
128×128×128, 80	~7.3×	93%
128×128×128, 10	~1.2×	37%
160×160×160, 120	~10×	95%

Numerics are bit-identical (torch.equal) at every size.

System: 12th Gen Intel Core i7-12800H (20 threads, OMP_NUM_THREADS=8); Linux 6.8.0-124-generic
x86_64, glibc 2.35; Python 3.10.12; PyTorch 2.12.1+cu130. Metric runs on CPU.

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• New tests added to cover the changes.
• The panoptic-quality metric and handler tests passed locally, along with black/isort/ruff/mypy.

[coderabbitai]

📝 Walkthrough

Walkthrough

_get_pairwise_iou now computes the IoU matrix with vectorized tensor operations, including early handling for empty instance sets and torch.bincount-based intersection and area calculations. Tests cover overlapping, disjoint, and empty ground-truth or prediction inputs.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly summarizes the main change: vectorizing _get_pairwise_iou with a joint bincount.
Description check	✅ Passed	The description includes a clear summary, benchmark data, and type/test status, matching the template well enough.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

monai/metrics/panoptic_quality.py (1)

259-264: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Document the contiguous-ID precondition.

The bincount reshape relies on IDs being contiguous from zero; sparse IDs can produce an oversized bincount and fail the reshape. Add a Google-style docstring describing inputs, output, and this precondition.

As per path instructions, “Docstrings should be present for all definition” using Google-style sections.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@monai/metrics/panoptic_quality.py` around lines 259 - 264, Document the
function containing the gt_flat/pred_flat intersection computation with a
Google-style docstring covering its inputs, returned output, and the requirement
that remap_instance_id produces contiguous IDs starting at zero for both ground
truth and predictions. Ensure the docstring follows the project’s requirement
that every definition has documentation.

Source: Path instructions

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@monai/metrics/panoptic_quality.py`:
- Around line 259-264: Document the function containing the gt_flat/pred_flat
intersection computation with a Google-style docstring covering its inputs,
returned output, and the requirement that remap_instance_id produces contiguous
IDs starting at zero for both ground truth and predictions. Ensure the docstring
follows the project’s requirement that every definition has documentation.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 65a7dd41-409c-436e-8620-91cb89109f33

📥 Commits

Reviewing files that changed from the base of the PR and between 5da2472 and f443085.

📒 Files selected for processing (2)

• monai/metrics/panoptic_quality.py
• tests/metrics/test_compute_panoptic_quality.py