VLCounter: Text-aware Visual Representation for Zero-Shot Object Counting

TL;DR

The paper addresses Zero-Shot Object Counting (ZSOC) by removing the need for exemplar discovery and avoiding error propagation in two-stage pipelines. It proposes an end-to-end framework, VLBase, built on CLIP, and its counting-tailored extension, VLCounter, which adds Semantic-conditioned Prompt Tuning (SPT), Learnable Affine Transformation (LAT), and Segment-aware Skip Connection (SaSC) to leverage multimodal representations for counting unseen classes. Empirical results on FSC147 and cross-dataset CARPK/PUCPR+ show state-of-the-art or competitive performance, with improved generalization and faster inference compared to exemplar-based approaches. The work demonstrates the practical impact of leveraging CLIP’s local-semantic patch representations and multi-level feature fusion to achieve robust zero-shot counting without bottlenecks from exemplar discovery.

Abstract

Zero-Shot Object Counting (ZSOC) aims to count referred instances of arbitrary classes in a query image without human-annotated exemplars. To deal with ZSOC, preceding studies proposed a two-stage pipeline: discovering exemplars and counting. However, there remains a challenge of vulnerability to error propagation of the sequentially designed two-stage process. In this work, an one-stage baseline, Visual-Language Baseline (VLBase), exploring the implicit association of the semantic-patch embeddings of CLIP is proposed. Subsequently, the extension of VLBase to Visual-language Counter (VLCounter) is achieved by incorporating three modules devised to tailor VLBase for object counting. First, Semantic-conditioned Prompt Tuning (SPT) is introduced within the image encoder to acquire target-highlighted representations. Second, Learnable Affine Transformation (LAT) is employed to translate the semantic-patch similarity map to be appropriate for the counting task. Lastly, the layer-wisely encoded features are transferred to the decoder through Segment-aware Skip Connection (SaSC) to keep the generalization capability for unseen classes. Through extensive experiments on FSC147, CARPK, and PUCPR+, the benefits of the end-to-end framework, VLCounter, are demonstrated.

Figures (9)

• Figure 1: Comparison between two-stage pipeline and one-stage pipeline (ours). The two-stage pipeline requires training the exemplar discoverer (orange) before the counter (blue), along with the need for an extra training dataset to optimize the discoverer. In contrast, our one-stage pipeline is designed to be simpler and does not necessitate any additional data or training stage.
• Figure 2: Overview of VLBase and VLCounter: each without and with colored components. The end-to-end baseline, VLBase, employs CLIP encoders to extract both image and text embeddings. Then, the decoder processes the image-text similarity map along with visual embeddings to count the number of specified objects. With three colored modules, VLCounter leverages the generalization capability of VLBase to be tailored for object counting.
• Figure 3: Illustration for Semantic-conditioned Prompt Tuning (SPT). In addition to learnable visual prompts (orange) in the image encoder, text features (yellow) are integrated to specify the desired semantics.
• Figure 4: The flow of Semantic-aware Skip Connection (SaSC) and architecture of feature projection block. Intermediate visual features are projected and filtered with an object-aware counting map $\hat{S}$ to produce object-relevant encoder features. Consequently, these are integrated into its counterpart in the decoder.
• Figure 5: Qualitative comparison of VLBase and VLCounter on the FSC-147 (Top 4 rows) and CARPK (Bottom 2 rows). Class names and counting values are shown at the right top of the query image ($I$) and the predicted density map, respectively.