Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CamReasoner: Reinforcing Camera Movement Understanding via Structured Spatial Reasoning

Hang Wu, Yujun Cai, Zehao Li, Haonan Ge, Bowen Sun, Junsong Yuan, Yiwei Wang

TL;DR

CamReasoner reframes camera movement understanding as structured ego-motion inference using the Observation-Thinking-Answer (O-T-A) paradigm, mitigating hallucinations from black-box cues by grounding reasoning in spatio-temporal geometry. It introduces a two-stage training pipeline (SFT followed by RL with GRPO and EMA-GRPO) and builds a Large-scale Inference Trajectory Suite (18k SFT, 38k RL) to teach cinematic logic and causal motion reasoning. Empirically, CamReasoner-7B achieves state-of-the-art binary accuracy (78.4%) and VQA accuracy (74.5%) on CameraBench, with notable strengths in confusable motions and robust reasoning across 9 skills and 81 sub-tasks. The work advances automated film analysis and controllable video synthesis by bridging perception, structured deduction, and 3D spatial reasoning.

Abstract

Understanding camera dynamics is a fundamental pillar of video spatial intelligence. However, existing multimodal models predominantly treat this task as a black-box classification, often confusing physically distinct motions by relying on superficial visual patterns rather than geometric cues. We present CamReasoner, a framework that reformulates camera movement understanding as a structured inference process to bridge the gap between perception and cinematic logic. Our approach centers on the Observation-Thinking-Answer (O-T-A) paradigm, which compels the model to decode spatio-temporal cues such as trajectories and view frustums within an explicit reasoning block. To instill this capability, we construct a Large-scale Inference Trajectory Suite comprising 18k SFT reasoning chains and 38k RL feedback samples. Notably, we are the first to employ RL for logical alignment in this domain, ensuring motion inferences are grounded in physical geometry rather than contextual guesswork. By applying Reinforcement Learning to the Observation-Think-Answer (O-T-A) reasoning paradigm, CamReasoner effectively suppresses hallucinations and achieves state-of-the-art performance across multiple benchmarks.

CamReasoner: Reinforcing Camera Movement Understanding via Structured Spatial Reasoning

TL;DR

CamReasoner reframes camera movement understanding as structured ego-motion inference using the Observation-Thinking-Answer (O-T-A) paradigm, mitigating hallucinations from black-box cues by grounding reasoning in spatio-temporal geometry. It introduces a two-stage training pipeline (SFT followed by RL with GRPO and EMA-GRPO) and builds a Large-scale Inference Trajectory Suite (18k SFT, 38k RL) to teach cinematic logic and causal motion reasoning. Empirically, CamReasoner-7B achieves state-of-the-art binary accuracy (78.4%) and VQA accuracy (74.5%) on CameraBench, with notable strengths in confusable motions and robust reasoning across 9 skills and 81 sub-tasks. The work advances automated film analysis and controllable video synthesis by bridging perception, structured deduction, and 3D spatial reasoning.

Abstract

Understanding camera dynamics is a fundamental pillar of video spatial intelligence. However, existing multimodal models predominantly treat this task as a black-box classification, often confusing physically distinct motions by relying on superficial visual patterns rather than geometric cues. We present CamReasoner, a framework that reformulates camera movement understanding as a structured inference process to bridge the gap between perception and cinematic logic. Our approach centers on the Observation-Thinking-Answer (O-T-A) paradigm, which compels the model to decode spatio-temporal cues such as trajectories and view frustums within an explicit reasoning block. To instill this capability, we construct a Large-scale Inference Trajectory Suite comprising 18k SFT reasoning chains and 38k RL feedback samples. Notably, we are the first to employ RL for logical alignment in this domain, ensuring motion inferences are grounded in physical geometry rather than contextual guesswork. By applying Reinforcement Learning to the Observation-Think-Answer (O-T-A) reasoning paradigm, CamReasoner effectively suppresses hallucinations and achieves state-of-the-art performance across multiple benchmarks.
Paper Structure (24 sections, 5 equations, 4 figures, 4 tables)

This paper contains 24 sections, 5 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Camera Movement Understanding
  4. Multimodal Reinforcement Learning
  5. Method
  6. Data Collection and Curation
  7. CamReasoning-SFT-18k Construction.
  8. CamReasoning-RL-38k
  9. Training Strategy
  10. Supervised Fine-tuning
  11. Reinforcement Learning Framework
  12. Problem Formulation.
  13. Reward Function Design.
  14. GRPO.
  15. Stability Strategy: EMA-GRPO.
  16. ...and 9 more sections

Figures (4)

  • Figure 1: Overview of CamReasoner. We propose the <observation>-<think>-<answer>paradigm to improve camera movement understanding. The figure illustrates our model generating detailed visual observations and logical thinking for movements like truck left and pan right.
  • Figure 2: Data generation pipeline for CamReasoning-SFT-18k. We utilize an answer-conditioned process to generate 38,672 initial trajectories, which are then filtered for format, accuracy, and consistency to retain 18,541 high-quality samples.
  • Figure 3: Qualitative results across four typical camera movements. For each case, we visualize the temporal frame sequence alongside the CamReasoner-7B response. The model demonstrates robust spatial reasoning by generating detailed <observation> of visual cues and a logical <think> process to accurately identify the movement and provide the final <answer>.
  • Figure 4: Training curves for SFT and RL phases. The top row shows the loss and grad_norm during the SFT process; the bottom row visualizes the convergence of accuracy, format, and overall rewards during RL training.