What

Selects cv2.INTER_AREA (and antialias=True on the PyTorch path) instead of the implicit bilinear default when the model's preprocessing resize is shrinking an image to the network input size. Applies to both resize modes:

• Stretch to — inference/core/models/roboflow.py
• Fit (black/white/grey edges) in — via resize_image_keeping_aspect_ratio in inference/core/utils/preprocess.py

Upscaling keeps bilinear (INTER_AREA degenerates toward nearest-neighbor when enlarging).

Why

The resize calls passed no interpolation argument, so OpenCV defaulted to INTER_LINEAR (bilinear). Bilinear samples only a 2×2 neighborhood per output pixel, so on large downscales (e.g. a 3120×4160 image → 640×640, ~6.5× per axis) it point-samples and aliases — high-frequency detail like thin, low-contrast embossed text is discarded rather than averaged in. INTER_AREA integrates over the entire source region each output pixel covers, acting as a proper anti-aliasing filter.

This surfaced on a customer (GE Vernova) serial-number OCR model: sending a full-resolution image silently dropped one of two detection regions, while resizing first with INTER_AREA (via the image_preprocessing workflow block, which already uses INTER_AREA) recovered both. The Models-page UI "worked" only incidentally — it downsamples to 1536×2048 before inference, reducing the final downscale ratio and thus the aliasing.

Impact / risk

• Detection coordinates are unaffected (output geometry unchanged; only sample quality improves).
• Better recall on high-resolution inputs with fine detail; no expected regression on inputs already near model size (downscale ratio ~1 → bilinear and area converge).
• Brings inference resize quality in line with the image_preprocessing block (INTER_AREA) and, likely, dataset-generation preprocessing — reducing train/serve skew. Open question for reviewers: please confirm the interpolation used during version generation so inference matches it exactly.

Repro (customer A/B, model serialnumber-001/15, same image)

Out of scope (follow-ups)

• The PyTorch antialias flag only applies to mode="bilinear"/"bicubic" — fine here, but worth a guard if modes are ever made configurable.
• Separate bug: Fit within, Fill (with center crop) in, and Fit (reflect edges) in silently fall back to Fit (black edges) in in roboflow.py.

Test plan

• Unit: assert INTER_AREA chosen when target < source and INTER_LINEAR when target ≥ source, for both Stretch to and the letterbox path.
• Regression: detection on a high-res fine-detail fixture recovers previously-missed boxes; near-model-size inputs unchanged within tolerance.

🤖 Generated with Claude Code