Counting Through Occlusion: Framework for Open World Amodal Counting

Abstract

Object counting has achieved remarkable success on visible instances, yet state-of-the-art (SOTA) methods fail under occlusion. This failure stems from a fundamental architectural limitation where backbone networks encode occluding surfaces rather than target objects, thereby corrupting the feature representations required for accurate enumeration. To address this, we present CountOCC, an amodal counting framework that explicitly reconstructs occluded object features through hierarchical multimodal guidance. Rather than accepting degraded encodings, we synthesize complete representations by integrating spatial context from visible fragments with semantic priors from text and visual embeddings, generating features at occluded locations across multiple pyramid levels. We further introduce a visual equivalence objective that enforces consistency in attention space, ensuring that both occluded and unoccluded views of the same scene produce spatially aligned gradient-based attention maps. Together, these complementary mechanisms preserve discriminative properties essential for accurate counting under occlusion. For rigorous evaluation, we establish occlusion-augmented versions of FSC-147 and CARPK (FSC-147-OCC and CARPK-OCC). CountOCC achieves SOTA performance on FSC-147-OCC with 26.72% and 20.80% MAE reduction over prior baselines under occlusion in validation and test, respectively. CountOCC also demonstrates exceptional generalization by setting new SOTA results on CARPK-OCC with 49.89% MAE reduction and on CAPTURe-Real with 28.79% MAE reduction, validating robust amodal counting.

The occlusion problem

Motivation

Method

Reconstruct, then align

CountOCC architecture with image and text encoders, Feature Reconstruction Module, VisEQ, feature enhancer and cross-modality decoder — **CountOCC architecture.** The Feature Reconstruction Module (FRM) replaces corrupted occluded tokens at each pyramid level; VisEQ aligns gradient-based attention maps of the occluded (student) and original (teacher) views. Reconstructed features flow through the feature enhancer and cross-modality decoder to produce visible and occluded counts.

Feature Reconstruction ModuleLearnable queries at occluded positions self-attend, cross-attend to visible tokens for spatial context, then cross-attend to fused text–exemplar embeddings for semantics, producing class-discriminative features where the occluder was.
Visual Equivalence (VisEQ)A teacher sees the clean image, a student sees the occluded one. Attention-similarity and ROI-consistency losses make their gradient-based attention maps agree, so localization does not depend on occlusion.
New benchmarksFSC-147-OCC and CARPK-OCC add controlled occlusion to standard open-world and car-counting datasets for rigorous amodal evaluation.

Results

FSC-147-OCC · MAE / RMSE · lower is better

Method	Prompt	Val MAE	Val RMSE	Test MAE	Test RMSE
CLIP-Count	Text	26.31	80.45	23.90	108.57
CounTX	Text	24.81	75.58	23.04	113.83
CounTR	Exemplars	23.14	66.78	22.25	104.75
LOCA	Exemplars	17.13	44.25	16.77	78.41
CountGD	Exemplars + Text	15.83	54.38	14.42	85.40
CountOCC	Exemplars + Text	11.60	35.40	11.42	38.68

Generalization

Benchmark	CountGD MAE	CountOCC MAE	Reduction
CARPK-OCC (test)	9.28	4.65	−49.9%
CAPTURe-Real	14.97	10.66	−28.8%

Qualitative comparison of CLIP-Count, CounTX, CounTR, LOCA, CountGD and CountOCC on occluded FSC-147 images — **Qualitative comparison on FSC-147-OCC.** Columns show CLIP-Count, CounTX, CounTR, LOCA, CountGD, and CountOCC. Prior methods undercount hidden objects; CountOCC counts correctly under occlusion across diverse categories.

Citation

@article{arib2025countocc,
  title   = {Counting Through Occlusion: Framework for Open World Amodal Counting},
  author  = {Arib, Safaeid Hossain and Akter, Rabeya and Chowdhury, Abdul Monaf and
             Sourov, Md Jubair Ahmed and Hasan, Md Mehedi},
  journal = {arXiv preprint arXiv:2511.12702},
  year    = {2025}
}