--- title: 'WATCH: Vision, Security & Smartwatch Research' url: https://www.emergentmind.com/topics/watch type: topic --- # WATCH: Vision, Security & Smartwatch Research Searching arXiv for recent papers and exact entries related to “WATCH” and associated watch/smartwatch/watch-time research. WATCH denotes several technically distinct constructs in recent arXiv literature. In computer vision, it refers to “World-aware Allied Trajectory and pose reconstruction for Camera and Human,” a monocular-video framework for jointly reconstructing camera and human motion in world coordinates [2509.04600]. In systems security, the name WATCH is used for a reverse-engineering, audit, and reimplementation effort around Apple Watch protocols, culminating in the Android-based WatchWitch platform [2507.07210]. Adjacent “watch” research also includes watch-time prediction in recommender systems [2401.07521], multimodal smartwatch sensing datasets [2512.03988], real-time multiwatch educational analytics [2512.02651], and behavioral inference from smartwatch motion traces [1709.09148]. This suggests that the term operates less as a single field label than as a cluster spanning vision, security, recommendation, and wearable sensing. ## 1. Terminological scope and research contexts The explicit acronym WATCH appears in at least two distinct forms. One is the computer-vision framework “World-aware Allied Trajectory and pose reconstruction for Camera and Human,” which addresses global human motion reconstruction from in-the-wild monocular videos by jointly modeling camera orientation, camera translation cues, and human motion [2509.04600]. The other is the Apple Watch interoperability and privacy effort described under the name WATCH and implemented as WatchWitch, which reverse-engineers wireless protocols, audits security, and provides an Android reimplementation with local-first privacy controls [2507.07210]. A broader watch-centered literature uses the word in non-acronymic but technically central ways. “Watch time” is treated as a long-tailed continuous target in video recommendation, motivating regression-by-classification, uncertainty modeling, and deconfounding objectives [2401.07521]. “Smartwatch” denotes a sensing platform for multimodal physiological acquisition, educational analytics, and affect inference from accelerometer traces [2512.03988]. “Watch” also names the act of observing action sequences, as in experiments on what makes gameplay clips enjoyable to watch [2605.30864]. The common thread is not lexical but methodological: each line studies either data generated by watches or signals derived from watching. ## 2. WATCH in global human motion reconstruction WATCH for monocular motion reconstruction takes as input video frames $\{I_t\}_{t=1}^T$ and, optionally, camera extrinsics $\{\mathbf R_t^{cam}, \mathbf t_t^{cam}\}_{t=1}^T$. Its feature stack combines ViT image features from a frozen HMR2.0 encoder, 2D keypoints from ViTPose, bounding-box features, camera rotation increments, and local camera velocity. These signals are passed through MLPs, summed into fused temporal features, and processed by a RoPE-augmented Transformer. A multi-task decoder predicts SMPL-X body parameters $(\theta_t,\beta_t)$, camera-space human root rotation and translation, camera roll-pitch, camera local velocity, and human local velocity in body coordinates [2509.04600]. The framework’s defining geometric step is analytical heading angle decomposition. It factorizes camera orientation as $$ \mathbf R_t^{cam} = \mathbf R_{yaw,t}^{cam}\,\mathbf R_{rp,t}^{cam}, $$ derives the pure heading increment by $$ \Delta\mathbf R_{yaw,t}^{cam} = \mathbf R_{rp,t}^{cam}\,\Delta\mathbf R_t^{cam}\,(\mathbf R_{rp,t+1}^{cam})^T, $$ and recursively integrates $$ \mathbf R_{yaw,t}^{cam} = \mathbf R_{yaw,0}^{cam}\;\prod_{i=1}^t\Delta\mathbf R_{yaw,i}^{cam}. $$ Human world-space orientation is then recovered as $$ \mathbf R_t^{h,w} = \mathbf R_{yaw,t}^{cam}\,\mathbf R_{rp,t}^{cam}\,\mathbf R_t^{h,c}. $$ Only roll and pitch must be estimated by the network; yaw is recovered analytically. The paper characterizes this as more efficient and extensible than prior geometric methods. WATCH also introduces a world-model-inspired camera trajectory integration mechanism. Instead of hard-decoding global human positions from SLAM poses, it treats camera velocity as spatial context. Predicted velocities are integrated by $$ \hat{\mathbf t}_t = \mathbf t_0 + \sum_{i=1}^t \mathbf R_i\,\hat{\mathbf v}_i, $$ applied separately to camera and human trajectories. Training uses a standard camera-space reconstruction loss together with human and camera trajectory consistency terms and a total loss $$ \mathcal L = \mathcal L_{hmr} + \lambda_h\,\mathcal L_{traj}^h + \lambda_{cam}\,\mathcal L_{traj}^{cam}. $$ The reported data sources are AMASS, BEDLAM, H36M, and 3DPW; optimization uses sequence length 120, batch size 128, and AdamW for approximately 500 epochs. On world-space benchmarks, WATCH reports WA-MPJPE$_{100}=74.3\text{mm}$ on RICH versus GVHMR’s $78.8\text{mm}$, and on EMDB-2 reports WA-MPJPE$_{100}=106.4$ versus $111.0$, RTE $=1.7\%$ versus $2.0\%$, jitter $=14.4$ versus $16.7$, and foot-sliding $=3.3$ versus $3.5$. In camera space, it reports 3DPW MPJPE $=54.5$ versus $55.6$ and PVE $=66.0$ versus $67.2$, while static-camera RICH slightly favors GVHMR in PA-MPJPE. Ablation on EMDB shows that trajectory integration lowers WA-MPJPE$_{100}$ by 3–4 mm and jitter by approximately 1–2 mm. The reported limitations are smaller gains in static-camera scenes and dependence on the quality of camera extrinsics or inertial input. ## 3. WATCH and WatchWitch for Apple Watch interoperability, security, and privacy The WATCH effort around Apple Watch combines technical reverse-engineering, a security and privacy audit, and an open-source Android reimplementation named WatchWitch [2507.07210]. At the transport layer, the reverse-engineered stack uses BLE ATT/GATT and L2CAP, with Apple multiplexing IP packets through a proprietary “Magnet/NRLP” shim. NRLP packets over L2CAP use the structure $$ \texttt{type}\,\|\,\texttt{length}_{4\text{ bytes}}\,\|\,\texttt{payload}_{\texttt{length}}\,\|\,\texttt{checksum}_{2\text{ bytes}}. $$ Once NRLP is established, IKEv2 over UDP/5000 and ESP are carried transparently. On Wi-Fi, the watch and phone switch automatically to IP/UDP combined with AES-GCM or ChaCha20-Poly1305 for ESP. Inside ESP frames, TCP ports 61314 and 61315 host an “NWSC” channel broker, above which Alloy provides a TLV-framed, proto-buf or bplist-wrapped message bus. For health data classified as “class A,” the system uses a custom “A-over-C” double-encryption scheme. The reported sequence is: choose random keys $k_1,k_2\in\{0,1\}^{128}$; compute $sed = \mathrm{AES\!-\!CBC}_{k_2}(p)$; compute $c_1 = \mathrm{AES\!-\!CTR}_{k_1}(k_2)$; compute $c_2 = \mathrm{RSA\!-\!OAEP}_{pk_r}(k_1\|\!c_1)$; sign with $s = \mathrm{ECDSA}_{sk_s}(c_2)$; and send $\{c_2,s,sed\}$ in an Alloy DataMessage. Reverse-engineering used palera1n and Dopamine for root access, Frida hooks, tcpdump, Apple PacketLogger, Ghidra, and IDA Pro. The security audit identifies three issues. First, unauthenticated IKEv2 Link-Director Notify payloads can be accepted in unencrypted IKE_INIT or rekey messages, allowing an attacker who can inject or modify BLE or Wi-Fi frames to redirect the watch to attacker-controlled Wi-Fi. Second, the outer A-over-C payload uses unauthenticated AES-CBC for the actual data, so an attacker with class C/D IPSec keys can exploit CBC malleability and introduce controlled bit-flips in decrypted plaintext. Third, deleting a cycle-tracking entry zeros out fields but leaves its deleted timestamp and sample type, permitting forensic inference about erased entries. WatchWitch is the Android reimplementation that demonstrates practical interoperability. Its high-level flow uses a jailbroken iPhone setup app to obtain long-term IPSec keys from the terminus daemon and transfer them to Android; then a custom IKEv2 Notify points the watch to the Android phone’s Wi-Fi IP; Android completes IKEv2 and builds class C and D ESP tunnels; class C traffic is routed to an Alloy server and class D to a Shoes proxy; and Alloy topics are dispatched to health sync, notification forwarding, and firewall modules. The Android-side components are KeyReceiver, IKEv2Handler, ESPInjector/Ejector, AlloyServer, ShoesServer, HealthSyncService, and NotificationService. The privacy architecture is explicitly local-first. A full copy of the watch’s Health SQLite database is maintained locally on Android, encrypted with SQLCipher, with the key stored in the TEE via Android Keystore. A user-configurable firewall tabulates Internet-bound requests by destination FQDN and Android package. The work defines Data Autonomy as $$ DA = 1 - \frac{\sum_{t\in \text{sent to cloud}} \text{size}(t)} {\sum_{t\in \text{all transactions}} \text{size}(t)}. $$ By blocking Apple cloud endpoints, the paper reports $DA=0.98$ for health sync, compared to approximately $0$ under stock watchOS. In lab measurements, WatchWitch sustained a stable watch-phone link for 24 hours and 2 hours of GPS-tracked workouts, with approximately 150 ms IKEv2 handshake delay, negligible ESP per-packet overhead, CPU load below 5%, and RAM footprint of 20 MB. The paper explicitly frames this as evidence that Apple’s interoperability limits are a choice rather than a technical necessity. ## 4. Watch-time prediction in video recommender systems A separate research line treats watch time as a significant indicator of user satisfaction in video recommender systems and formulates its prediction under a sharply right-skewed and long-tailed continuous target distribution [2401.07521]. In short-video feeds, approximately 30% of impressions end within 3 s and 80% within 32 s. Direct $L_1$ or $L_2$ regression under-represents rare large-watch-time events and ignores ordinal ranking relations. CREAD addresses this with a Classification-Restoration framework with Error-Adaptive Discretization. It defines thresholds $0=t_0t_m)$ for $m=1,\dots,M-1$, and restores a scalar estimate via the identity $$ E[T|x]=\int_0^{T_{\max}} P[T>t|x]\,dt, $$ discretized as $$ \hat y(x)=\sum_{m=1}^M \hat\phi_m(x)\cdot(t_m-t_{m-1}). $$ Its central theoretical claim is an error decomposition $$ E(\hat y-T)^2 = V_p + V_w + V_b + V_y, $$ with variance from estimating bucket probabilities, estimating bucket means, bucket-mean replacement bias, and irreducible within-bucket variance. Error-Adaptive Discretization minimizes a combined upper bound $$ D^* = \arg\min_D \{A_w(D) + \beta\,A_b(D)\}, $$ balancing learning error and restoration error. The full loss is $$ L=\lambda_{ce}\,L_{ce} + \lambda_{restore}\,L_{restore} + \lambda_{ord}\,L_{ord}. $$ Offline results report KuaiRec XAUC improving from 0.5677 to 0.6009 and MAE from 3.3090 to 3.2150; CIKM XAUC from 0.6568 to 0.6671 and MAE from 2.5499 to 1.7928; and industrial MAE from 19.0740 to 18.6716. An online A/B test on Kwai App with 20% traffic reports a statistically significant +0.29% lift in average watch time, with no adverse effect on like, share, or follow metrics. TPM, the Tree-based Progressive Regression Model, decomposes watch time into a series of conditionally dependent classification tasks arranged in a tree [2306.03392]. With cut-points $0=r_0r_i\mid y>r_{i-1},x), $$ so that $$ P(y>r_i\mid x)=\prod_{j=1}^i p_j(x). $$ Expected watch time can be written as $$ \mathbb E[y\mid x] = \sum_{i=1}^K \Delta r_i\;\prod_{j=1}^i p_j(x), $$ and predictive variance is used as an uncertainty regularizer. TPM also incorporates backdoor adjustment by conditioning on a confounder $D$ and averaging over $P(d)$ at inference. On Kuaishou, it reports MAE 4.741 and XAUC 0.599 versus D2Q’s 5.426 and 0.565; on CIKM16, MAE 0.884 and XAUC 0.676. Online deployment in Kuaishou with 300 million DAUs and 20% test traffic reports consistent daily watch-time lifts between +0.210% and +0.265%. EXUM extends watch-time prediction with explicit uncertainty modeling [2504.07575]. It introduces a confidence predictor $c_i=g_\phi(x_i)\in[0,1]$ and ensembles model outputs with labels during training. In the quantile-prediction case, $$ p'_i = c_i\,p_i + (1-c_i)\,y_i, $$ and in ordinal regression, $$ p'_{i,n}=c_i\,p_{i,n} + (1-c_i)\,y_{i,n}. $$ To avoid collapse to $c_i=0$, it adds an adversarial confidence-maximization term $$ \mathcal L_c = -\sum_i \ln c_i. $$ The framework is deployed around CREAD in the reranking stage of an industrial short-video platform, while dropping the confidence head at inference so that there is zero extra latency in production. The reported online result is a statistically significant +0.31% watch-time gain, with likes, follows, and comments changing by less than 0.3%. Taken together, these systems show a strong convergence toward ordinal or distributional modeling, explicit treatment of uncertainty, and production validation through A/B testing. A plausible implication is that watch-time prediction has become a benchmark problem for integrating statistical calibration, ranking fidelity, and systems constraints. ## 5. Smartwatch sensing datasets and multimodal infrastructure HEART-Watch is a multimodal physiological dataset collected from a Google Pixel Watch 2 across sitting, standing, and walking, with synchronized chest ECG reference and intermittent blood pressure measurements [2512.03988]. The cohort consists of 40 healthy adults: 23 female and 17 male, age 19–75 years with mean $\pm$ SD $=44.2\pm21.0$ years. The race/ethnicity breakdown spans eight categories, and body-type diversity is described by height and weight ranges. Each physical state includes 4 min continuous recording, and between states there are 5 intermittent upper-arm blood pressure measurements using Omron BP5000. The smartwatch modalities are wrist ECG at approximately 250 Hz, two-channel PPG at approximately 68.4 Hz, and 3-axis accelerometer at approximately 6.27 Hz, all timestamped in Unix epoch with nanosecond precision and stored via Android Room before CSV extraction. Synchronization uses zero-centering of ECG and PPG, linear interpolation to 250 Hz, coarse timestamp alignment, and fine alignment by a constant shift estimated from mean ECG R-peak differences. The dataset provides both native-rate raw CSVs and synchronized 250 Hz multimodal CSVs without filters. The paper also specifies preprocessing and feature definitions used for analyses but not applied to the published CSVs. It defines instantaneous heart rate from ECG as $$ IHR_i = 60/(RR_i-RR_{i-1}), $$ and HRV measures $$ SDNN = \sqrt{\frac{1}{N-1}\sum_1^N (IBI_i-\bar{IBI})^2},\qquad RMSSD = \sqrt{\frac{1}{N-1}\sum_1^N (IBI_{i+1}-IBI_i)^2}. $$ Pulse transit time is $$ PTT = t_{\text{foot of PPG waveform}} - t_{\text{ECG R-peak}}. $$ Signal-quality results report 0% weak or invalid 30 s segments for chest ECG, PPG, and ACC, and smartwatch ECG weak/invalid rates of 5.3% in sitting, 5.0% in standing, and 12.5% in walking. Bland–Altman analysis versus chest ECG yields QRS duration bias of +1.51 ms, +1.04 ms, and +1.93 ms across sitting, standing, and walking, while PR interval bias is +10.09 ms, +13.22 ms, and +17.71 ms. Access is controlled via UHN OneDrive and requires a study plan and data-use agreement. Watch-DMLT and ViSeDOPS extend smartwatch sensing into synchronized, multi-user educational analytics [2512.02651]. Watch-DMLT is built on the Fitbit SDK with three modules: sensor data collection on the Fitbit Sense 2, data transmission on a paired smartphone, and data processing and storage on a secure server. Sensor streams and nominal sampling rates are heart rate at 1 Hz, accelerometer at 25 Hz, gyroscope at 25 Hz, and quaternion orientation at 25 Hz, totaling approximately 251 samples/s and approximately 2–3 kB/s per device. CSV rows include UTC timestamp, device identifier, and sensor readings, and residual drift is corrected by a linear model $$ \Delta t(t)= (t_{\mathrm{ref}}-t_{\mathrm{dev}})-\alpha\,(t_{\mathrm{dev}}-t_0), $$ with corrected time $$ t_{\mathrm{corrected}} = t_{\mathrm{dev}}+\Delta t(t_{\mathrm{dev}}). $$ ViSeDOPS uses Python, Flask, Pandas, NumPy/SciPy, Plotly Dash, OpenCV, CSV logs, MP4 video, and JSON annotations to merge synchronized multimodal streams. Derived metrics include rolling mean heart rate, RMSSD-based HRV, acceleration-magnitude variance, and a multimodal engagement index $$ E(t)=w_1\,\widetilde{\overline{\mathrm{HR}(t)}}+w_2\,[1-\widetilde{\sigma}_{\mathrm{acc}}(t)]+w_3\,\widetilde{F_{\mathrm{gaze}}(t)}. $$ In a classroom deployment with 65 students and up to 16 watches, the system collected approximately 700 minutes of multimodal data. Mean presenter heart rate was $80.2\pm12.1$ bpm, mean RMSSD was $42.5\pm10.7$ ms, and mean motion variance was $0.56\,\mathrm{m}^2/\mathrm{s}^4$ with gesturing spikes up to $2.3\,\mathrm{m}^2/\mathrm{s}^4$. Cross-modal correlations were $r=0.31\pm0.12$ for HR versus motion energy, $r=-0.28\pm0.09$ for HRV versus gaze fixation rate, and $r=0.45\pm0.15$ for engagement index versus slide transitions. The system handled up to 16 watches with approximately 2% end-to-end sample loss, clock drift below 100 ms over 2 hours after NTP sync, sub-200 ms alignment accuracy, and dashboard latency below 2 s. ## 6. Behavioral and perceptual inference from watch-associated signals Smartwatch accelerometry has been used to infer emotion from gait [1709.09148]. In a preliminary study with 50 young adults, participants experienced both happy and sad priming, were assigned either audio-visual or audio-only stimuli, and then walked a 250 m S-shaped corridor wearing a Samsung Gear S2 on the left wrist and a Polar H7 chest strap. The reported analysis uses only accelerometer data. After a moving average of length three, $$ \bar x[n]=(x[n-1]+x[n]+x[n+1])/3, $$ the signal is divided into consecutive, non-overlapping one-second windows of approximately 23 samples. From each window, 17 time-domain features are extracted, including mean, maximum, minimum, standard deviation, energy, kurtosis, skewness, RMS, RSS, AUC, absolute area, mean of absolute values, range, quartiles, MAD, an angle-based orientation feature, and vector magnitude average. For each of 44 participants after exclusion of 6 with missing data, personal models were trained with logistic regression and random forest under five-fold cross-validation. Per-participant accuracies range from approximately 50% up to approximately 85%, with the bulk between 60% and 80%, and both classifiers substantially outperform the 50% majority baseline. The paper identifies the window-independence assumption, absence of frequency-domain features, and lack of generalized cross-subject models as limitations. A distinct line studies the enjoyment of action sequences that people watch rather than signals produced by a wrist-worn watch [2605.30864]. Using 24 Flappy Bird-style gameplay clips, the study independently manipulates task difficulty and moment-to-moment uncertainty, termed dangerousness. Difficulty is defined from simulated completion failure rates over 1,000 trials for each of three agents, $$ Difficulty_j = 1-\frac{1}{3}\sum_{i=1}^3 SR_i^{(j)}, $$ and dangerousness is defined from the negative value function, $$ Danger(s_t)=-V(s_t),\qquad Danger_{\text{trajectory}}=\frac{1}{T}\sum_{t=1}^T[-V(s_t)]. $$ Among 905 U.S. adults recruited via Prolific, 41 were excluded, leaving $N=864$, with 288 assigned to each of the Enjoyment, Apparent Difficulty, and Apparent Dangerousness rating conditions. Inter-rater reliability after Spearman–Brown correction is $0.57$ for enjoyment, $0.83$ for difficulty, and $0.93$ for dangerousness. Mean dangerousness ratings correlate with model-based dangerousness at $0.79$, and mean difficulty ratings correlate with model-based difficulty at $0.64$. Enjoyment correlates with perceived difficulty at $0.68$ but not with perceived dangerousness, where the reported correlation is $0.11$ with $p=0.62$. In an OLS model, difficulty predicts enjoyment with $R^2=0.44$, while dangerousness does not improve fit. The paper therefore argues that viewers derive pleasure primarily from watching agents surmount challenging environments rather than from narrow escapes alone. Across these studies, watch-associated inference covers both first-person sensing and third-person observation. This suggests a broad methodological continuum: accelerometers, ECG, PPG, gaze, and video can all serve as observational substrates, but the inferential target may be internal state, engagement, or the structure of the watched event itself. Source: https://www.emergentmind.com/topics/watch