Add masked reconstruction branch to inference loop with entropy loss and geometric consistency

[Copilot]

Implements geometry-aligned masked image modeling (MIM) in the inference path for test-time adaptation, with entropy minimization for geometric consistency.

Changes

Model extensions (rtdetr_pose/model.py):

• RenderTeacher: Small CNN encoding mask+depth → geometry features
• DecoderMIM: Reconstructs masked P5 neck features
• RTDETRPose.forward() accepts geom_input (mask+normalized depth), feature_mask, return_mim flag
• Automatic spatial resizing for mask/geometry inputs

Loss functions (rtdetr_pose/losses.py):

• mim_reconstruction_loss(): L1 loss on masked regions between reconstructed and teacher features
• entropy_loss(): Minimizes prediction entropy for confident geometric alignment
• Losses class integrates both with configurable weights (0.1, 0.01)

API:

# Opt-in via flag (disabled by default)
model = RTDETRPose(enable_mim=True)

# Geometry input: [mask, log(depth/z_ref)]
geom = create_geom_input(mask, depth_obj)
mask_feat = generate_block_mask(h, w, mask_prob=0.6)

# Returns standard outputs + optional MIM branch
outputs = model(image, geom_input=geom, feature_mask=mask_feat, return_mim=True)

# MIM outputs: recon_feat, teacher_feat, entropy
loss = mim_reconstruction_loss(outputs["mim"]["recon_feat"], 
                                outputs["mim"]["teacher_feat"],
                                mask=outputs["mim"]["mask"])

Test-time training:
Adapt normalization parameters via MIM + entropy loss on unlabeled test images with predicted geometry.

Compatibility:
Zero overhead when enable_mim=False (default). Existing inference paths unchanged.

Implementation notes

• Operates on P5 features (1/16 to 1/32 scale)
• Teacher features use stopgrad() for stable supervision
• Reference depth normalization: log(D_obj / median(D_obj))
• ~100K additional parameters (RenderTeacher + DecoderMIM)

Implements RT-DETR 6DoF Spec v0.4 §6 (Geometry-Aligned MIM).

Original prompt

inference loop に masked reconstruction branc

entropy loss + geom consistency

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[Copilot]

Copilot wasn't able to review any files in this pull request.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.