Context Aware Grounded Teacher
for Source-Free Object Detection

A bias-aware SFOD framework that grounds the teacher through relational context modeling, semantic augmentation, and expert foundational supervision, achieving new state-of-the-art on natural and medical imaging benchmarks.

🧠  Accepted in International Journal of Computer Vision
Tajamul Ashraf1,3*
Rajes Manna3†
Partha Sarathi Purkayastha4†
Tavaheed Tariq3
Janibul Bashir3*
1 Department of Computer Vision, MBZUAI, Abu Dhabi, UAE 3 GAASH Research Lab, NIT Srinagar, J&K, India 4 Microsoft Research India, Bengaluru, India
* Corresponding author  ·  † Equal contribution
arXiv PDF Code GAASH Research Lab

Abstract

Source-free object detection (SFOD) faces persistent challenges due to class imbalance–driven context bias and instability in teacher–student training under noisy pseudo-labels. Existing techniques tend to ignore the context bias and class imbalance shift, especially in medical data. To tackle this, we propose Grounded Teacher (GT), a bias-aware source-free framework that grounds the teacher model through relational and semantic regularization. GT introduces a Relational Context Module (RCM), which maintains an EMA estimate of cross-domain contextual bias, modeling directional inter-class confusions. Building upon this, a Semantic Augmentation (SA) strategy selectively augments minority and confusable classes through adaptive MixUp in both source-similar and source-dissimilar target regions. A Semantic-Aware Loss (SAL) applies diagonally normalized weights, preventing gradient explosion while emphasizing minority–majority corrections. Additionally, a frozen Expert branch derived from large vision foundation models (LVFMs) serves as a supervisory reference during training, refining pseudo-label quality without adding inference overhead. Evaluations on Cityscapes→Foggy (50.8 mAP) and medical transfers (+5.9 AP50 on DDSM→INBreast) show consistent gains and improved minority-class detection, with less than 12% additional training cost.

Why SFOD Remains Hard

Three compounding problems existing methods fail to address simultaneously

⚠️

Mode Collapse via Pseudo-label Noise

Biased teacher models produce skewed pseudo-labels. As EMA propagates student weights back, error accumulates and the teacher collapses—predicting only dominant classes.

🔀

Severe Domain Shift

Source-free setting prohibits access to source data during adaptation. Models must transfer knowledge blindly, compounding context bias under large source-target gaps.

⚖️

Class Imbalance in Medical Data

Breast-cancer datasets like DDSM have severe imbalance; most images lack majority-class instances. Conventional image-level resampling is ineffective for detection.

How Grounded Teacher Works

Three interlocking modules that break the bias-collapse cycle in source-free settings

MODULE 01

Relational Context Module (RCM)

Maintains an EMA-updated global confusion matrix across training. Models directional inter-class confusions so downstream augmentation and loss focus on genuinely problematic pairs.

MODULE 02

Semantic Augmentation (SA) + Semantic-Aware Loss (SAL)

RCM-guided relation-MixUp blends minority instances with their most-confused majority neighbours from a FIFO Cropbank. SAL applies diagonal-normalized, bounded weights to classification loss—preventing gradient explosion while amplifying minority corrections.

MODULE 03

Expert Foundational Branch

A frozen Large Vision Foundation Model (LVFM — BioMedParse for medical, GroundingDINO for natural) supervises the student via pseudo-label regression + SAL. Training-only: zero inference overhead.

EMA update Teacher ← Student weights Pseudo-label feedback Filtered by τ = 0.8 confidence Expert supervision Training-only · no inference cost Domain split Variance-based source-similar / dissimilar

Framework Architecture

The GT framework consists of a student–teacher pair connected by EMA weight updates. The Relational Context Module sits between them, maintaining a running confusion matrix that guides both the Cropbank sampling strategy and the Semantic-Aware Loss.

A frozen Expert Branch (BioMedParse / GroundingDINO) provides an additional supervision signal to the student during training. At inference, only the student runs, the expert adds zero latency.

The variance-based domain splitter partitions target images into source-similar and source-dissimilar subsets, allowing tailored augmentation strength in each region.

Grounded Teacher Architecture Diagram

State-of-the-Art Performance

Evaluated on natural and medical domain adaptation benchmarks

Table 4 — Cityscapes → Foggy Cityscapes
Method SF Venue Person Rider Car Truck Bus Train M.cycle Bicycle mAP
Source Only ✗ — 22.4 26.6 28.5 9.0 16.0 4.3 15.2 25.3 18.4
DA-Faster ✗ CVPR'18 29.2 40.4 43.4 19.7 38.3 28.5 23.7 32.7 32.0
EPM ✗ ECCV'20 44.2 46.6 58.5 24.8 45.2 29.1 28.6 34.6 39.0
AT ✗ CVPR'22 43.7 54.1 62.3 31.9 54.4 49.3 35.2 47.9 47.4
MRT ✗ ICCV'23 52.8 51.7 68.7 35.9 58.1 54.5 41.0 47.1 51.2
SFOD-Mosaic ✓ AAAI'21 25.5 44.5 40.7 33.2 22.2 28.4 34.1 39.0 33.5
HCL ✓ NeurIPS'21 38.7 46.0 47.9 33.0 45.7 38.9 32.8 34.9 39.7
LODS ✓ CVPR'22 34.0 45.7 48.8 27.3 39.7 19.6 33.2 37.8 35.8
IRG ✓ CVPR'23 37.4 45.2 51.9 24.4 39.6 25.2 31.5 41.6 37.1
PETS ✓ ICCV'23 46.1 52.8 63.4 21.8 46.7 5.5 37.4 48.4 40.3
SF-UT ✓ ECCV'24 40.9 48.0 58.9 29.6 51.9 50.2 36.2 44.1 45.0
LPLD ✓ ECCV'24 39.7 49.1 56.6 29.6 46.3 26.4 36.1 43.6 40.9
OursGT ✓ IJCV'26 42.7 55.4 61.7 40.7 62.0 54.6 39.1 53.0 50.8
Oracle ✗ — 66.3 61.1 80.8 45.6 68.8 52.0 49.1 54.9 59.8
Medical Imaging Benchmarks (FROC Recall & F1)

Table 1 — DDSM → INBreast (large-to-small, digitised film → full-field digital)

Method SF Venue R@0.05 R@0.30 R@0.50 R@1.00 AUC F1
IRG ✓ CVPR'23 0.05 0.05 0.07 0.09 0.110 0.120
LPLD ✓ ECCV'24 0.09 0.15 0.35 0.35 0.548 0.635
D-MASTER ✗ MICCAI'24 0.25 0.61 0.70 0.82 0.808 0.524
OursGT ✓ IJCV'26 0.06 0.45 0.65 0.92 0.589 0.758

Table 2 — DDSM → RSNA-BSD1K (cross-geography, US acquisition)

Method SF Venue R@0.05 R@0.30 R@0.50 R@1.00 AUC F1
IRG ✓ CVPR'23 0.16 0.25 0.37 0.39 0.308 0.235
MRT ✗ ICCV'23 0.32 0.52 0.69 0.72 0.741 0.352
OursGT ✓ IJCV'26 0.10 0.43 0.58 0.91 0.781 0.530

Table 3 — RSNA → INBreast (cross-machine acquisition)

Method SF Venue R@0.05 R@0.30 R@0.50 R@1.00 AUC F1
IRG ✓ CVPR'23 0.05 0.05 0.07 0.09 0.110 0.120
MRT ✗ ICCV'23 0.03 0.09 0.12 0.17 0.739 0.587
OursGT ✓ IJCV'26 0.01 0.28 0.49 0.90 0.638 0.605

Qualitative Results

Visual comparisons of GT against prior SFOD methods across natural and medical domain transfers. Detection boxes are colour-coded:

  • Green — True Positive
  • Blue — Misclassified
  • Red — False Negative
  • Pink — False Positive

GT consistently recovers minority-class detections, trucks and buses in fog, and lesions in mammograms, that source-free baselines miss entirely.

Foggy Cityscapes qualitative results
DDSM to INBreast qualitative results
DDSM to RSNA qualitative results
RSNA to INBreast qualitative results
Foggy Cityscapes — GT vs MRT / IRG / LPLD. GT recovers trucks and buses missed by baselines.