--- title: 'PerSense: Dual Domains in Personality & Vision' url: https://www.emergentmind.com/topics/persense type: topic --- # PerSense: Dual Domains in Personality & Vision Searching arXiv for relevant "PerSense" papers and adjacent work to ground the entry. arxiv_search(query="PerSense", max_results=10, sort_by="relevance") arxiv_search(query="\"PerSense\" personalized instance segmentation", max_results=10, sort_by="relevance") arxiv_search(query="\"PerSense\" personality assessment text commonsense reasoning", max_results=10, sort_by="relevance") PerSense is an ambiguous research name on arXiv rather than a single canonical system. In one usage, it denotes a 2020 framework for estimating human personality traits from text and using those traits for machine commonsense reasoning [2004.09275]. In another, it denotes a 2024 computer-vision framework for training-free, model-agnostic, one-shot personalized instance segmentation in dense images, later extended as PerSense++ in 2025 [2405.13518], [2508.14660]. The shared label obscures the fact that these works belong to unrelated technical lineages—one in NLP and computational personality modeling, the other in dense-scene segmentation and prompt-based vision systems. ## 1. Name, scope, and disambiguation The term “PerSense” has been used for distinct systems with different problem formulations, data regimes, and evaluation protocols. The most direct way to interpret the name is therefore bibliographically rather than conceptually. | Usage | Domain | Core purpose | |---|---|---| | "Personality Assessment from Text for Machine Commonsense Reasoning" [2004.09275] | NLP / computational personality | Estimate personality traits from text and use them to predict responses to open-ended commonsense questions | | "PerSense: Personalized Instance Segmentation in Dense Images" [2405.13518] | Computer vision | Segment all instances of a target category in crowded scenes from one support example | | "Towards PerSense++: Advancing Training-Free Personalized Instance Segmentation in Dense Images" [2508.14660] | Computer vision | Extend PerSense with stronger exemplar selection, hybrid prompt generation, and mask rejection | The name should not be conflated with "PerSE" for personalized evaluation of open-ended text generation [2310.03304], "PerSEval" for assessing personalization in text summarizers [2407.00453], or "SentiPers," a Persian sentiment-analysis corpus [1801.07737]. Those are separate systems with different objectives and abbreviatory logic. ## 2. PerSense as personality assessment from text The 2020 PerSense framework starts from the hypothesis that commonsense reasoning is shaped not only by facts or logic but also by human traits, preferences, biases, and values. Its objective is threefold: estimate personality traits from naturally produced text, use those traits to support commonsense reasoning, and do so with lower computational complexity than large proprietary systems such as IBM Watson [2004.09275]. The framework is organized as a three-phase pipeline. First, text is collected from Amazon movie reviews, stream-of-consciousness essays, and WordPress blog posts. Linguistic features are then extracted, especially adjectives and their frequencies. These texts are labeled by IBM Watson Personality Insights, which outputs personality scores in the interval \([0,1]\). Second, PerSense predicts personality by either a probabilistic aggregated-PDF method or a supervised ML method. Third, the inferred personality information is used for a commonsense prediction task in which the input consists of answers to a 50-question Big Five questionnaire and the output is a predicted answer to a subjective commonsense question. The work is framed around the Big Five dimensions—Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism—but the reported experiments emphasize Neuroticism. The stated reason is that Neuroticism had the most balanced score distribution in the corpus, whereas traits such as Openness were heavily concentrated near one end of the scale. The paper also motivates this focus by linking Neuroticism to mental health analysis and suicide prevention. This formulation places PerSense within a specific strand of computational psychometrics: passive personality profiling from digital text rather than questionnaire-only assessment. Its commonsense component is likewise narrower than general commonsense QA. The target is prediction of subjective human answers in settings where multiple responses are plausible, including policy opinions and “Family Feud”-style prompts. ## 3. Probabilistic and ML formulations in the 2020 system The aggregated-PDF method models how informative words distribute across personality scores. For a trait \(T\) and word \(w\), the system estimates a density \(f_{w,T}(s)\) over \(s \in [0,1]\). For a new text containing informative words \(w_1,\dots,w_m\), these densities are combined as $$ \Phi_T(s)=\frac{\prod \limits_{i=1}^mf_{w_i,T}(s)}{\int_{x=0}^1\prod \limits_{i=1}^mf_{w_i,T}(x)dx}. $$ The predicted personality score is the peak of \(\Phi_T(s)\). The method assumes conditional independence among informative words. The paper describes a confidence factor as the logarithmic ratio of the first to the second maximum probability. A natural mathematical expression for that description is $$ \text{Confidence} = \log\left(\frac{p_1}{p_2}\right), $$ where \(p_1\) and \(p_2\) are the highest and second-highest peaks, respectively. The ML branch uses adjective-frequency vectors as inputs and IBM Watson-derived scores as labels. The reported model family includes Support Vector Regression with RBF kernel, Linear Regression, Random Forest Regression, SVMs with linear, polynomial, and RBF kernels, Decision Tree, K-Nearest Neighbors, Perceptron, Multi-layer Perceptron, and Random Forest Classifier. The summary reports a filtered ML dataset with **16,934 data points**, **345 input features**, and **8 class labels** per personality trait in classification; the Neuroticism subsection separately states that the corpus was reduced from **17,128** to **16,834** after filtering. This suggests an internal reporting mismatch in the paper summary rather than a single stable dataset count. For Neuroticism, the PDF approach reports **MAE = 15.5\%** and **RMSE = 19.5\%**, with **MAE = 10.5\%** for samples whose confidence is greater than 2. The abstract further states that the PDF approach reaches **97\% accuracy** when the confidence factor is greater than 3. In the ML comparison, the **Multi-layer Perceptron** is reported as the strongest model, with **82.2\% accuracy** in the abstract; the conclusion separately reports **88.2\%**, which again indicates a reporting inconsistency. For the commonsense task, a **Random Forest Classifier** is reported to achieve **82.3\% accuracy**, and one example improves from **52\% before fusion to 76\% after fusion** when semantically similar answer labels are merged [2004.09275]. These details are important because they delimit what PerSense actually demonstrates. It does not establish personality assessment from text in a clinical sense, and it does not solve generic commonsense reasoning. Rather, it shows that personality-linked features can be exploited to predict certain subjective responses under a supervised learning setup. ## 4. PerSense as personalized instance segmentation in dense images The 2024 PerSense paper reuses the same name for a wholly different problem: segmenting all instances of a support-specified object category in crowded scenes without training on the target category [2405.13518]. The setting is one-shot and support-query based. Inputs are a query image \(I_Q\), a support image \(I_S\), and a support mask \(M_S\). The support image is masked, a class label is extracted by a VLM using the prompt “Name the object in the image?”, and that label is then used to guide grounding in the query image. A central design choice is to convert dense-scene localization into a density-map-guided prompt-generation problem. The query-support similarity is written as $$ S_{\text{score}(Q, S_{\text{supp})=\cos\_sim(f(Q), f(S_{\text{support}))}, $$ and the most similar pixel in the highest-confidence grounded box is used as a positive prior: $$ P_{\max}=\arg\max_{P\in B_{\max} S_{\text{score}(P,S_{\text{supp})}. $$ A density map generator then produces a density map for the query image. PerSense does not treat the density map as a mask. Instead, it uses an **Instance Detection Module (IDM)** to turn the density map into candidate point prompts and a **Point Prompt Selection Module (PPSM)** to filter false positives before sending the surviving prompts to SAM. The IDM operates by grayscale conversion, thresholding, erosion, contour extraction, and composite-contour handling. The thresholded image is defined as $$ I_{\text{binary}(x,y)= \begin{cases} 1 & \text{if } I_{\text{gray}(x,y)\ge T \\ 0 & \text{if } I_{\text{gray}(x,y) \mu + 2\sigma. $$ For composite regions, PerSense applies a distance transform and thresholds at \(0.5 \cdot \max(\text{dist\_transform})\) to recover child contours. The PPSM then filters candidate points by similarity and spatial consistency. Its adaptive threshold is defined as $$ \text{sim\_threshold} = \frac{\text{max\_score}}{\text{object\_count}/\sqrt{2}}, $$ and a candidate is kept only if its similarity exceeds this threshold and it lies inside a grounded detection. A feedback mechanism closes the loop: top candidates from the initial segmentation are reused as exemplars to improve density-map generation, with **\(k=4\)** selected in the reported setting. The overall framework is described as **training-free**, **model-agnostic**, and **one-shot** because it composes pretrained CLE, grounding, density-map generation, and SAM modules entirely at inference time. ## 5. PerSense-D, PerSense++, and empirical results The 2024 paper introduces **PerSense-D** as a dedicated benchmark for personalized instance segmentation in dense images, while the 2025 follow-up extends both the method and the benchmark description [2405.13518], [2508.14660]. The earlier paper reports **717 images total**, comprising **689 dense query images** and **28 support images**, across **28 object categories**, with **28,395** total objects, an average of **39 objects per image**, a **7 / 218** min/max object count per image, and average resolution **839 × 967**. The 2025 paper reports the same **717 images** and **28 object categories** but gives **36,837 annotated objects** and an average of **53 objects per image**. This suggests the benchmark accounting was revised or expanded between versions. On PerSense-D, the 2024 paper reports the following segmentation results: | Method | mIoU | Avg time/image | |---|---:|---:| | PerSAM | 24.45 | 39.78 s | | PerSAM-F | 29.34 | 47.78 s-ish (39.78 + 8) | | Matcher | 62.78 | 10.2 s | | Grounded-SAM | 65.92 | 1.8 s | | PerSense (DSALVANet) | 70.96 | 2.7 s | | PerSense (CounTR) | 71.61 | 2.7 s | The same paper reports component-wise gains from PPSM and the feedback mechanism. For **DSALVANet**, the baseline is **65.58**, **+ PPSM = 66.95**, and **+ feedback = 70.96**. For **CounTR**, the baseline is **68.12**, **+ PPSM = 70.58**, and **+ feedback = 71.61**. The best PPSM normalization factor is reported as \(\sqrt{2}\), and the best GroundingDINO box threshold is **0.15**. The 2025 PerSense++ paper adds three components: **diversity-aware exemplar selection**, **hybrid IDM**, and **Irrelevant Mask Rejection Module (IMRM)**. Diversity-aware selection combines semantic closeness and geometric consistency through a weighted score $$ \Phi(b_i) = w_1 S_{\text{cent}, i} + w_2 S_{\text{ref}, i} - w_3 \log(\Delta A_i) - w_4 \Delta A_{d, i}, $$ and enforces scale diversity through percentile-based small, medium, and large bins. The hybrid IDM supplements contour-based prompts with density peaks, while IMRM rejects spatially inconsistent masks using IQR-based outlier analysis plus a majority-cluster estimate, with an IoU-based safeguard retaining masks whose overlap with grounded detection boxes is at least **0.8**. PerSense++ improves the reported PerSense numbers substantially. On PerSense-D, the paper reports **70.96 mIoU** for PerSense (DMG1) versus **77.45 mIoU** for PerSense++ (DMG1), and **71.61 mIoU** for PerSense (DMG2) versus **81.35 mIoU** for PerSense++ (DMG2). The same work also evaluates dense subsets **COCO-20\(^d\)** and **LVIS-92\(^d\)** and states that the improvements are larger on dense subsets than on the full datasets [2508.14660]. ## 6. Conceptual differences, limitations, and common misconceptions The two principal arXiv usages of PerSense are unrelated in method and scientific aim. The 2020 system is a computational personality framework that links adjective-based text features, IBM Watson trait scores, and supervised prediction of subjective commonsense responses. The 2024–2025 system is a prompt-generation and segmentation pipeline that combines grounding, density estimation, and SAM-based decoding in crowded scenes. Treating them as parts of a single research program would therefore be incorrect. Their limitations are also domain-specific. For the 2020 PerSense, the personality ground truth comes from **IBM Watson Personality Insights** rather than standard clinical questionnaires; the deep empirical analysis is concentrated on **Neuroticism**; the commonsense dataset is modest; and the reported summary itself contains several number mismatches, including **16,934** versus **16,834** filtered data points and **82.2\%** versus **88.2\%** MLP accuracy [2004.09275]. For the vision PerSense line, performance depends on density-map quality, the method is explicitly designed for dense rather than sparse scenes, strong intra-class variation can hurt one-shot localization, and composite-contour handling remains heuristic [2405.13518]. A plausible commonality is that both systems operationalize “personalization” from limited conditioning information: the NLP version conditions on a person’s text or questionnaire profile, while the vision version conditions on a single support exemplar. That similarity is only nominal and structural, not methodological. In current arXiv usage, “PerSense” therefore denotes an overloaded label spanning at least two separate research objects: one in personality-aware NLP and one in dense-image personalized segmentation, with the latter continuing as PerSense++ [2508.14660]. Source: https://www.emergentmind.com/topics/persense