fix(tflite): match PyTorch predict() preprocessing and mask decoder

[omkar-334]

What does this PR do?

The TFLite inference helper produced inputs the PyTorch graph never saw, and decoded masks with a different half-pixel convention than PostProcess.forward. After this change, FP32 TFLite is bit-exact to PyTorch for both object detection and segmentation. FP16 and dynamic-range int8 are within IoU 0.965 at production thresholds with zero class mismatches.

The conversion graph itself was never broken: a same-input probe shows boxes/logits max|diff| <= 0.0001 between PyTorch and the existing FP32 .tflite. Everything we were seeing came from the inference-side pre/post-processing being slightly off from the training-time pipeline.

Changes

Two helpers added to inference.py to fix where the TFLite predict path diverges from PyTorch's.

_preprocess_image (extracted from _run_inference). The old code called PIL.Image.resize with no filter argument. PIL's default has been BICUBIC since Pillow 9.1, but RFDETR.predict resizes with BILINEAR via torchvision.transforms.functional.resize. Same image, different input tensor, model sees something it wasn't trained on. The new helper uses torchvision.transforms.functional (to_tensor → resize → normalize) when torch is importable for byte-for-byte parity, and falls back to PIL.Image.resize with explicit BILINEAR for torch-free deployments (still close to PyTorch, ~0.016 max residual).

_decode_masks (rewritten) + _bilinear_resize_half_pixel (new fallback). The old _decode_masks did PIL.Image.fromarray(..., "F").resize(..., BILINEAR) per mask. PIL's BILINEAR uses a corner-aligned half-pixel convention; PyTorch's F.interpolate(..., mode="bilinear", align_corners=False) (what PostProcess.forward calls) uses centered half-pixel. The shift is sub-pixel, but on thin or border-hugging masks it flipped enough pixels to drop mask IoU to ~0.5 even when the source logits matched. The rewrite calls torch.nn.functional.interpolate directly when torch is importable, and falls back to _bilinear_resize_half_pixel — a vectorised numpy implementation of centered half-pixel bilinear — when it isn't.

Before vs after (apples-to-apples at threshold 0.3)

Object detection (RFDETRNano on dog/people-walking/bus/zidane, 43 detections total):

Mode	Metric	Before	After	Note
FP32	IoU_worst	0.640	1.000	the dog image was the regression source
FP32	conf_max_diff	0.129	0.000
FP32	class_mismatches	1	0
FP16	IoU_worst	0.406	0.996	the IoU_worst FP16 disaster
FP16	conf_max_diff	0.167	0.012
FP16	class_mismatches	2	0
dyn_int8	IoU_mean	0.979	0.99	small lift; int8 was already weight-only

Related to #1024

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Testing

• I have tested this change locally
• I have added/updated tests for this change

[codecov]

Codecov Report

❌ Patch coverage is 94.91525% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 81%. Comparing base (04200cf) to head (91cdb5d).
⚠️ Report is 5 commits behind head on develop.

❌ Your project check has failed because the head coverage (81%) is below the target coverage (95%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #1131   +/-   ##
=======================================
  Coverage       81%     81%           
=======================================
  Files          108     108           
  Lines        11226   11269   +43     
=======================================
+ Hits          9144    9184   +40     
- Misses        2082    2085    +3     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR fixes divergence between the TFLite inference helper and the PyTorch RFDETR.predict() / PostProcess.forward pipeline by aligning image preprocessing (resize + ImageNet normalization) and segmentation mask upsampling (half-pixel bilinear, align_corners=False) so TFLite outputs match PyTorch much more closely.

Changes:

• Adds _preprocess_image() to standardize TFLite input preprocessing and avoid Pillow’s default resize filter mismatch.
• Rewrites _decode_masks() to use a centered half-pixel bilinear convention via a new _bilinear_resize_half_pixel() helper.
• Updates _run_inference() to use the new preprocessing path and pass NHWC batched input directly to the interpreter.