ProTracker: Probabilistic Integration for Robust and Accurate Point Tracking Tingyang Zhang, Chen Wang, Zhiyang Dou, Qingzhe Gao, Jiahui Lei, Baoquan Chen, Lingjie Liu Abstract We propose ProTracker , a novel framework for accurate and robust long-term dense tracking of arbitrary points in videos. Previous methods relying on global cost volumes effectively handle large occlusions and scene changes but lack precision and temporal awareness. In contrast, local iteration-based methods accurately track smoothly transforming scenes but face challenges with occlusions and drift. To address these issues, we propose a probabilistic framework that marries the strengths of both paradigms by leveraging local optical flow for predictions and refined global heatmaps for observations. This design effectively combines global semantic information with temporally aware low-level features, enabling precise and robust long-term tracking of arbitrary points in videos. Extensive experiments demonstrate that ProTracker attains state-of-the-art performance among optimization-based approaches and surpasses supervised feed-forward methods on multiple benchmarks. The code and model will be released after publication. Figure 1 : Visualization of tracking trajectories in various videos.Our method achieves robust point tracking without suffering from drifting over time, even in challenging scenarios such as occlusions and multiple similar regions. 1 Introduction Point tracking models [ 42 , 40 , 29 , 41 , 12 , 23 ] provide critical motion and deformation cues in scenes, thus they are essential for video analysis, especially for tasks like 4D reconstruction [ 25 , 44 , 51 ] and video editing [ 16 ] . The recent focus of point tracking is long-term dense tracking of any pixel in a video, also known as Tracking Any Point (TAP) [ 12 ] . Existing methods can be broadly classified into two categories. 1) Supervised tracking models [ 17 , 23 , 52 , 13 , 11 , 10 , 27 , 12 ] . Specifically, TAP-net [ 12 ] predicts trajectories by generating heatmaps that capture the relationship between the target point and the rest of the frames, while some others [ 17 , 23 , 52 , 27 ] iteratively refine the trajectory of the same point within a temporal window. These supervised learning-based trackers have achieved promising results on existing benchmarks, but they often struggle to generalize to out-of-domain inputs, as they are typically trained on specific datasets. Some of them either disregard temporal information [ 12 ] or suffer from context drift and loss particularly during extended occlusions as they rely on sliding window techniques [ 23 , 52 , 17 , 27 ] . 2) Optimization based models [ 50 , 43 , 28 , 49 ] . Based on test-time optimization, they have gained attention by leveraging the priors in foundation models trained on web-scale datasets. For instance, some methods [ 50 , 43 , 28 ] represent the entire scene as a quasi-3D canonical volume and use 3D bijections to map local coordinates to a global 3D canonical space, allowing for consistent tracking of points. However, the proxy canonical space represented by neural networks tends to be overly smooth, which limits tracking accuracy. DINO-Tracker [ 49 ] fine-tunes a feature extractor and heatmap refiner using the strong semantic priors from DINOv2 to track through long-term occlusions. However, challenges arise when the features are not distinct enough or when multiple similar parts are present in the scene. In this paper, we present ProTracker for accurate and robust point tracking. The key idea of our method is a bidirectional Probabilistic Integration for both optical flow predictions and long-term correspondences, inspired by Kalman Filter [ 22 ] . Specifically, we begin with removing incorrect initial predictions to reduce their negative impact on subsequent estimations with a hybrid filter including an object-level filter [ 37 ] and a geometry-aware feature filter [ 53 ] . For the remaining rough optical flow predictions, we address the inherent noise in optical flow estimates by introducing a probabilistic integration method that treats each prediction as a Gaussian distribution and merges them into a single Gaussian distribution to identify the most likely point prediction. The integration is done in both forward and backward directions for highly accurate and robust flow estimation. However, optical flow is limited to visible objects and tends to fail when a point disappears and then reappears in a different location, resulting in missing segments in the trajectory. To improve performance in challenging long-term point tracking as well as the occlusion problem, we train a long-term feature correspondence model and use it to identify keypoint positions across frames with discriminative features. Then, we jointly integrate flow estimation and long-term keypoints to obtain the final prediction. This combination equips the model to robustly recover trajectory segments and mitigate drift during long-term tracking. We conduct extensive experiments to evaluate our method on TAP-Vid benchmarks. Among optimization-based approaches, our method surpasses all previous methods across all metrics. Additionally, it demonstrates competitive performance even when compared to supervised feed-forward methods and achieves the highest accuracy in position estimation among all approaches. In summary, our contributions are as follows: • We propose ProTracker, a novel probabilistic integration framework that merges multiple rough predictions and significantly enhances the accuracy and robustness of point tracking. • We incorporate long-term correspondence matching into our probabilistic integration framework to address both long-term tracking and occlusion, enabling precise point tracking over extended durations. • Our method achieves the state-of-the-art performance among test-time training approaches while demonstrating competitive results compared to data-driven methods. 2 Related Work Optical flow aims to establish dense motion estimations between consecutive frames. Classical methods [ 19 , 30 , 4 , 3 ] optimize warp field with smoothness as regularization. Modern data-driven methods [ 21 , 15 , 20 , 45 , 47 ] learn deep neural networks to generate or refine flow predictions based on large amounts of annotated data, which has significantly improved performance. Although optical flow methods can accurately predict displacements between adjacent frames, they often fail when the displacement is too large due to biases in the training data, tending to keep points stationary. This makes optical flow unsuitable for direct long-term tracking. Even chaining flow predictions across frames can lead to drift and other issues. In our approach, we use RAFT [ 47 ] as the primary tracking tool, with the aid of additional models to perform precise point tracking. Dense correspondence involves finding pixel-level matches between an arbitrary pair of images. Correlation volumes are constructed to measure the similarity between pairs of pixels based on classic [ 29 ] or learning-based [ 9 , 32 , 38 , 48 ] feature descriptor, and the accurate point matches are decided accordingly. Recently, large pretrained visual foundation models [ 36 , 6 , 34 , 39 ] have shown their ability to extract powerful features and can be combined for robust matching across different scene/object appearances [ 18 , 31 , 46 , 53 , 8 ] . While directly using these correspondences for point tracking lacks accuracy [ 1 , 49 ] due to the lower resolution of the features compared to the original image, they can effectively serve as a filtering tool to discard incorrect predictions. Figure 2 : Pipeline overview of our proposed method. (1) Sample & Chain: Key points are initially sampled and linked through optical flow chaining to produce preliminary trajectory predictions. (2) Long-term Correspondence: Key points are re-localized over longer time spans to maintain continuity, even for points that temporarily disappear. (3) Hybrid Filter: Masks and feature filters are applied to remove incorrect predictions, reducing noise for subsequent steps. (4) Probabilistic Integration: Filtered flow predictions across frames are first integrated and then combined with long-term keypoint to produce the final prediction, producing smoother and more consistent trajectories. Tracking any point aims to track arbitrary points across a whole video, recovering the full trajectory and occlusion state. TAP-net [ 12 ] directly predicts via finding the target in a refined heatmap. PIPs [ 17 ] proposes to iteratively refine the trajectory within a temporal window according to the spatial context. Many attempts have been made to improve the refinement process. Co-tracker [ 23 ] counted in the relation between points and designed a self-attention to support them with each other. SpatialTracker [ 52 ] lifts points to 3d space and performs tracking with spatially meaningful information. TAPTR [ 27 ] treats points as queries and updates them in a DETR [ 5 ] style. Some other methods like TAPIR [ 13 ] and LocoTrack [ 11 ] adopt a coarse-to-fine strategy, dividing the tracking process into initialization and iterative optimization phases, which allows the well-initialized points to guide the trajectory in other frames. While those supervised methods may be limited to their spatial or temporal field of view due to large memory cost, Omnimotion [ 50 ] first proposes to learn a 3d representation for each video with color and pre-computed optical flow as self-supervision, in which a bijective mapping enables the query of any point in a different frame. Decomotion [ 28 ] decomposes the scene representation into static and dynamic and utilizes a temporal invariant feature as extra supervision. CaDeX++ [ 43 ] leverages a depth estimator to speed up and a more efficient deformation network. DINO-Tracker [ 49 ] trains a delta feature extractor as compensation for the powerful DINO [ 34 ] feature. MFT [ 33 ] is a zero-shot method that directly chains optical flow and selects the most reliable estimation as the final tracking prediction. However, problems like drift may occur when facing long videos. 3 Method Given an image sequence \{{I^{t}}\}_{t=1}^{T} from a monocular video, our goal is to take a query pixel \bm{p}^{t}\in\mathbb{R}^{2} from an arbitrary frame I^{t} as input and predict its trajectories \{\hat{\bm{p}}^{t}\}_{t=1}^{T} over the video, along with the occlusion prediction \{\hat{o}^{t}\}_{t=1}^{T} , which is known as the TAP (Tracking Any Point) problem. As shown in Fig. 2 , our pipeline first obtains both initial rough optical flow predictions from multiple previous frames and long-term correspondence predictions. We filter unreliable point predictions with an object-level segmentation model [ 37 ] and geometry-aware semantic features [ 53 ] (Sec. 3.1 ). Then, we integrate multiple optical flow predictions in a probabilistic manner to get an integrated prediction from flow (Sec. 3.2 ). We further use a joint probabilistic integration between optical flow prediction and long-term semantic correspondence prediction to aggregate both global and local contexts, prevent drift and allow re-localization after reappearance (Sec. 3.3 ). Figure 3 : Bidirectional Probabilistic Flow Integration. Top row: Optical flow effectively tracks a point in the short term but may fail under occlusion due to its local nature. Bottom row: Long-term correspondence aids in globally relocating the target when the tracked point reappears. Once relocation is achieved, optical flow can resume tracking in the surrounding frames. 3.1 Hybrid Filter Since our method relies on rough predictions from optical flow and long-term correspondence for probabilistic integration (as shown in Fig. 2 ), inaccurate rough predictions can lead to cumulative errors and distort entire trajectories, which may significantly degrade tracking accuracy. To mitigate these issues, we propose a hybrid filter to abandon these predictions and avoid using them in the following probabilistic integration. Our hybrid filter consists of an object-level filter and a geometry-aware feature filter. First, an object-level segmentation model [ 37 ] generates masks associated with target points, filtering out predictions outside relevant objects and using global context to mitigate the impact of outliers; see ablation study in Fig. 5 . This step significantly benefits optical flow-based systems like RAFT [ 47 ] , which often struggle with occlusions due to their reliance on local feature matching. To further reduce ambiguity between semantically similar points and prevent flickering across different regions, an additional geometry-aware feature extractor [ 53 ] is employed. For each point, if its feature similarity to the original query point falls below 0.5, the point is classified as occluded, ensuring that only reliable predictions are retained and preventing errors from propagating due to semantic ambiguity. Together, the object-level and geometry-aware filters consist of a module which can distinguish different points from a global semantic perspective, thereby avoiding confusion between different objects or similar regions caused by local inductive bias from optical flow. 3.2 Bidirectional Probabilistic Flow Integration We introduce a probabilistic integration strategy inspired by the Kalman filter [ 22 ] , enabling frame-by-frame trajectory recovery from both semantics and low-level features. Our method incorporates both forward and backward passes, leveraging forward and backward flows to reconstruct complete trajectories. Further details are provided below. Forward Integration Our method sequentially predicts point trajectory and occlusion on the current frame based on previous predictions. Since every estimate is subject to noise, we extend both the track predictions and the optical flow into two-dimensional Gaussian distributions. Thus, we denote the predictions of frame i as (\bm{\mu}_{i},\bm{\Sigma}_{i}) , where \bm{\mu}_{i} and \bm{\Sigma}_{i} are the mean and covariance of the Gaussian distribution. We further assume these Gaussian distributions are isotropic and simplify the covariance matrix as \bm{\Sigma}_{i}=\sigma_{i}^{2}\bm{I} , where \bm{I} is the identity matrix. For the initial frame, we assume zero uncertainty ( \sigma_{0}=0 ). For any frame i>0 , we first calculate the flow chain estimations based on previous frames \{j_{1},j_{2},...,j_{n}\} . Given the prediction for frame j\in\{j_{1},j_{2},...,j_{n}\}\>(j(j>i) , denoted as (\bm{\mu}_{j},\sigma_{j}) , and the optical flow from frame j to frame i , (f_{ji},\sigma_{ji}) , we can obtain the mean and variance of the prediction for frame i after the same filtering process: \bm{\mu}_{i}^{b}=\frac{\sum_{j}\bm{\mu}_{ji}/\sigma_{ji}^{2}}{\sum_{j}1/\sigma_{ji}^{2}},\quad\sigma_{i}^{b}=\sqrt{\frac{(N-1)\cdot p+1}{\sum_{j}1/\sigma_{ji}^{2}}}. (7) If a point marked as occluded in the forward pass is visible in the backward pass, we adopt the backward result instead. Otherwise, we retain the forward prediction. This ensures less visible point ignored by the tracker, particularly in cases of occlusion. The backward pass helps recover points that are difficult to track from earlier frames but can be more easily tracked from later ones. Method DAVIS-First DAVIS-Strided Kinetics-First BADJA \delta_{avg}^{x}\uparrow OA \uparrow AJ \uparrow \delta_{avg}^{x}\uparrow OA \uparrow AJ \uparrow \delta_{avg}^{x}\uparrow OA \uparrow AJ \uparrow \delta^{seg}\uparrow \delta^{3px}\uparrow Omnimotion [ 50 ] - - - 67.5 85.3 51.7 - - - 45.2 6.9 MFT [ 33 ] 66.8 77.8 47.3 70.8 86.9 56.1 60.8 75.6 39.4 50.8 7.2 CaDeX++ [ 43 ] - - - 77.4 85.9 59.4 - - - - - DecoMotion [ 28 ] 69.9 84.2 53.0 74.4 87.2 60.2 - - - - - DINOTracker [ 49 ] 74.9 86.4 58.3 78.2 87.5 62.3 69.5 86.3 55.5 72.4 14.3 Ours 77.6 87.3 62.0 80.8 88.7 65.3 71.1 89.6 56.7 73.3 14.4 Table 1 : We compare our method with state-of-the-art optimization-based trackers on the TAP-Vid and BADJA benchmarks. We include MFT, which directly integrates optical flow for tracking. Our method consistently achieves superior performance across all metrics, with the best results bolded . Method DAVIS-First DAVIS-Strided Kinetics-First BADJA \delta_{avg}^{x}\uparrow OA \uparrow AJ \uparrow \delta_{avg}^{x}\uparrow OA \uparrow AJ \uparrow \delta_{avg}^{x}\uparrow OA \uparrow AJ \uparrow \delta^{seg}\uparrow \delta^{3px}\uparrow TAP-Net [ 12 ] 48.6 78.8 33.0 53.4 81.4 38.4 56.3 83.6 42.7 45.4 9.6 PIPs [ 17 ] 64.8 77.7 42.2 59.4 82.1 42.0 47.6 78.5 31.1 59.6 9.4 TAPIR [ 13 ] 70.0 86.5 56.2 74.7 89.4 62.8 63.6 86.4 52.6 68.7 10.5 CoTracker [ 23 ] 75.4 89.3 60.6 79.2 89.3 65.1 65.9 88.0 52.8 64.0 11.2 SpatialTracker [ 52 ] 76.3 89.5 61.1 - - - 67.1 88.3 53.9 63.6 10.6 BootsTAPIR [ 14 ] 74.0 88.4 61.4 78.5 90.7 66.4 - - - 72.6 13.4 CoTracker3 [ 24 ] 76.9 91.2 64.4 - - - 70.9 86.9 57.9 71.4 11.1 TAPTRv2 [ 26 ] 75.9 91.4 63.5 78.8 91.3 66.4 64.4 85.7 50.8 70.0 8.4 LocoTrack [ 11 ] 75.3 87.2 63.0 79.6 89.9 67.8 68.8 87.5 56.0 70.2 9.9 Ours 77.6 87.3 62.0 80.8 88.7 65.3 71.1 89.6 56.7 73.3 14.4 Table 2 : We compare our method with supervised feed-forward trackers on the TAP-Vid and BADJA benchmarks, where our method achieves the highest \delta_{avg}^{x} across all datasets and produces competitive results in OA and AJ. The best results are bolded . Figure 4 : We evaluate our method against state-of-the-art approaches, including feed-forward models (Co-Tracker3 [ 24 ] , SpatialTracker [ 52 ] , LocoTrack-B [ 11 ] , TAPTRv2 [ 27 ] ) and test-time training methods (CaDex++ [ 43 ] , DINO-Tracker [ 49 ] ). Experiments are conducted on the bike-packing and goat scenes from TAPVid-DAVIS, with tracking at 256×256 resolution and visualizations at the original resolution. More qualitative results are provided in the supplementary material. 3.3 Joint Flow and Long-term Correspondence Integration While flow integration can partially mitigate drift and produce smooth trajectories, accumulated errors still lead to drift over longer time spans. Moreover, in cases where an object disappears and reappears after some time, the optical flow method may struggle to track the point. To tackle these issues, we propose to integrate long-term correspondence into our flow-based prediction framework. We train a feature extractor \mathbf{\Phi}_{\Delta} and heatmap refiner \bm{\mathcal{R}} of a long-term correspondence-based keypoint tracker based on DINO-Tracker [ 49 ] for the input video, with the optical flow as a self-supervised signal. For frame i , the feature map \bm{\mathcal{F}}_{i} can be calculated as: \bm{\mathcal{F}}_{i}=\mathbf{\Phi}_{\text{DINO}}(I^{i})+\mathbf{\Phi}_{\Delta}(I^{i}) (8) After getting the query feature \bm{f}_{\text{query}} by sampling on the query point in \bm{p}_{0} , long-term predictions \bm{p}_{i} are generated by applying SoftArgMax on the refined heatmap: \bm{p}_{i}=\text{SoftArgMax}(\bm{\mathcal{R}}(\bm{f}_{\text{query}}\cdot\bm{\mathcal{F}}_{i})) (9) To avoid the negative impact of incorrect correspondences, we only require high-confidence keypoints. Thus, we only select those points with a cosine similarity greater than a threshold as keypoints. We incorporate these points into our probabilistic integration framework, as described by the following equation: \bm{\mu}_{i}^{\text{key}}=\bm{p}_{i},\text{if }\bm{\mathcal{F}}_{i}(\bm{p}_{i})\cdot\bm{f}_{\text{query}}>\rho,\sigma_{i}^{\text{key}}=1\\ (10) where \rho=0.7 , same as DINO-Tracker [ 49 ] . Specifically, whenever valid keypoints from the long-term correspondence are available, we treat them as another source of noisy observations besides optical flow. In this way, we can jointly integrate the flow prediction and long-term keypoint in our probabilistic integration framework (Sec. 3.2 ) to yield a final optimal estimation of the point’s location. Formally, let \bm{\mu}_{i} and \sigma_{i} represent the mean and variance from flow integration, and \bm{\mu}_{i}^{\text{key}} and \sigma_{i}^{\text{key}} represent the mean and variance from the key point observations. Then the combined estimates \bm{\mu}_{i}^{\text{final}} and \sigma_{i}^{\text{final}} are computed as: \bm{\mu}_{i}^{\text{final}}=\frac{\bm{\mu}_{i}/\sigma_{i}^{2}+\bm{\mu}_{i}^{\text{key}}/(\sigma_{i}^{\text{key}})^{2}}{1/\sigma_{i}^{2}+1/(\sigma_{i}^{\text{key}})^{2}}, (11) \sigma_{i}^{\text{final}}=\frac{1}{1/\sigma_{i}^{2}+1/(\sigma_{i}^{\text{key}})^{2}}. (12) The final prediction for the point location is given by \hat{\bm{p}}^{t}=\bm{\mu}_{i}^{\text{final}} , which indicates that our method’s result has maximum likelihood within the final prediction distribution. If neither the optical flow nor the long-term key point is valid in the current frame, the point is marked as occluded . By incorporating long-term key points, our approach mitigates drift, effectively aligning flow estimation with long-term keypoints to maintain trajectory accuracy. Moreover, it enables the model to re-localize points that have temporarily disappeared and reappeared in different locations and can track through sudden scene transitions. As a result, the flow-based predictions continue to refine the point’s trajectory, ensuring accurate and smooth tracking across frames, as demonstrated in Fig. 3 . 4 Experiments 4.1 Experiment Setup Dataset: We evaluate our method on the following datasets from TAP-Vid [ 12 ] and BADJA [ 2 ] : • TAPVid-DAVIS , a real-world dataset comprising 30 256px videos from DAVIS 2017 [ 35 ] . Each video contains between 34 and 104 RGB frames, capturing both camera movements and dynamic scene motions. We employ both the query-first mode, where all query points come from the first frame, and the query-strided mode, where a query is performed every 5 frames, for evaluation on DAVIS. • TAPVid-Kinetics , includes 1,189 videos, each with 250 frames at 256px resolution from Kinetics-700-2020 [ 7 ] . The dataset predominantly focuses on human activity, with both camera and object motion. Since our method includes test-time optimization steps, we use the subset of 100 videos sampled by Omnimotion [ 50 ] and the query-first mode for evaluation. • BADJA , consists of nine videos, at 480px resolution, showcasing animal movements in nature, with ground truth information for keypoint locations. Metrics: In accordance with the TAP-Vid [ 12 ] benchmark, we use the following metrics: • \delta_{avg}^{x} measures the percentage of visible points that are tracked within a specific pixel error from the ground truth, which is evaluated over five thresholds: {1,2,4,8,16} pixels, with the final score being the average fraction of points within these distances. • Occlusion Accuracy (OA) measures the fraction of points with correct visibility predictions in each frame, including both visible and occluded points. • Average Jaccard (AJ) measures both position and occlusion accuracy based on \delta^{x}_{avg} thresholds, which assesses the ratio of correctly predicted visible points to false predicted points. For evaluating BADJA [ 2 ] ,the following metrics are used: • \delta^{seg} measures the percentage of predicted points that lie within the distance of 0.2\sqrt{A} from the ground-truth position, where A is the area of the object mask. • \delta^{3px} measure the accuracy within a threshold of 3px. 4.2 Comparisons Baselines We compare our ProTracker to state-of-the-art methods: 1) optimization-based trackers such as Omnimotion [ 50 ] , CaDeX++ [ 43 ] , DecoMotion [ 28 ] and DINO-Tracker [ 49 ] . Note that optimization-based methods do not require model training on labeled datasets. 2) feedforward trackers trained in a supervised manner , including PIPs [ 17 ] , TAP-Net [ 12 ] , TAPIR [ 13 ] , CoTracker [ 23 ] , SpatialTracker [ 52 ] , BootsTAP [ 14 ] , CoTracker3 [ 24 ] , LocoTrack-B [ 11 ] and TAPTRv2 [ 26 ] . We additionally incorporate MFT [ 33 ] , which leverages RAFT to directly obtain trajectory predictions. Quantitative comparisons Tab. 1 and Tab. 2 compare our method against state-of-the-art trackers on the TAP-Vid and BADJA benchmarks, where our approach achieves the highest \delta_{avg}^{x} across all datasets, demonstrating superior precision in tracking visible points. We attribute this to our probabilistic model, which enhances tracking accuracy by integrating optical flow with long-term correspondence, resulting in smoother and more precise trajectory segments. Additionally, in terms of Occlusion Accuracy (OA) and Average Jaccard (AJ), our method performs on par with the best approaches, effectively handling occlusions while maintaining geometric consistency. Notably, on BADJA, our method outperforms all other trackers. Besides, among tracking methods that require test-time training, our method achieves the best performance across all metrics, demonstrating its robustness in both position and occlusion tracking. Additionally, our approach does not predict occlusion via trajectory agreement, making it a more efficient approach than DINO-Tracker [ 49 ] ; refer to Appendix 4 for more discussion. Qualitative results In Fig. 4 , our method could consistently trace the target points, even in challenging scenarios where objects frequently disappear and reappear. For instance, in the bike-packing sequence, our tracking points are placed on the person. While some methods (CoTracker3, SpatialTracker, TAPTRv2, CaDeX++) occasionally mis-track these points onto the bike or background, or fail to capture finer details such as the hands (DINO-Tracker, LocoTrack), our approach maintains precise tracking throughout the video. In the goat sequence, where the hooves frequently cross and obscure each other, our method reliably tracks the target points, remaining unaffected by overlapping limbs. We refer readers to our supplementary materials for more visual results. Our method successfully locates the positions of most visible points at each time step through seamless integration of key components. The hybrid filter prevents drift during occlusions (e.g., the bike in bike-packing, overlapping legs in goat), while long-term keypoints enable reliable re-identification. Our probabilistic framework anchors these keypoints, using temporal-aware optical flow to track less distinctive points, ensuring each point is located to the greatest extent possible. 4.3 Ablation study Next, we show an ablation study on different components of our framework using TapVid-DAVIS. Specifically, w/o keypoint removes joint integration with long-term key points, using only flow integration results. w/o geo-aware omits filtering by the geometry-aware feature. w/o mask applies rough flow predictions without object-level filtering. w/o probabilistic replaces probabilistic integration with selecting the prediction of the lowest \sigma as the final result. As shown in Tab. 3 and Fig. 5 , mask filtering effectively removes incorrect flow predictions, significantly improving precision and occlusion handling. Long-term key points enable trajectory recovery beyond optical flow, leading to substantial performance gains. Probabilistic integration further enhances positional accuracy by reducing uncertainty. Additionally, the geometry-aware feature mitigates misalignment in visually similar regions, improving occlusion handling accuracy. Method DAVIS-First DAVIS-Strided \delta_{avg}^{x} OA AJ \delta_{avg}^{x} OA AJ w/o key point 71.2 79.2 47.8 74.6 87.6 62.8 w/o geo-aware 77.6 85.7 60.0 80.8 88.4 63.3 w/o mask 72.3 82.3 57.1 73.9 82.6 57.8 w/o probabilistic 76.9 87.2 61.3 79.0 88.2 63.9 Ours Full 77.6 87.3 62.0 80.8 88.7 65.3 Table 3 : Ablation on different components on TAP-Vid-DAVIS. The best results are bolded . Figure 5 : Ablation study on different components. 5 Conclusion and Future Work In this paper, we introduced a robust tracking framework that combines optical flow integration with long-term correspondence through probabilistic integration to achieve accurate and smooth point tracking in dynamic video sequences. By incorporating object-level filtering, bidirectional probabilistic integration, and geometry-aware feature extraction, our method effectively mitigates drift, handles occlusions, and re-localizes temporally disappearing points. Our method outperforms traditional methods in handling complex motions and extended time gaps, demonstrating the advantages of integrating short-term and long-term information for reliable tracking. While our method provides robust tracking, its reliance on test-time training for keypoint extraction reduces efficiency compared to supervised approaches—a common limitation of optimization-based tracking methods. This dependency on test-time training arises due to the current feature extractor’s insufficient resolution and lack of temporal awareness. Future improvements in high-resolution feature extraction could help avoid test-time training and improve differentiation between objects and regions, allowing for fully unsupervised and real-time dense tracking. Supplementary Material 1 Video Results Please refer to our Supplementary Webpage for the corresponding videos of images illustrated in the paper and more results on different data. 2 Implementation Details 2.1 Hyperparameters During the dual filtering stage, we apply different thresholds to predictions from flow and long-term keypoints. For long-term keypoints, we only need those with higher confidence to avoid mistakes. A prediction is first marked as invalid if the cosine similarity to the query point on the refined DINO feature falls below 0.7, following DINO-Tracker. It is then filtered following the same procedure by geometry-aware feature with the similarity threshold of 0.5, which is a common practice. For predictions from flow, however, we want them to help track areas with less distinct features. Thus we use a threshold of 0.3 instead. These distinct thresholds allow flow to track featureless areas while ensuring that long-term keypoints do not drift into visually similar regions. Note that we use the same hyperparameters for all videos and our method don’t require any hyperparameter tuning. 2.2 Flow Preparation We utilize the RAFT [ 47 ] model, as adopted by MFT [ 33 ] , as our flow estimation model. It takes two images, \bm{I}_{j} and \bm{I}_{i} , captured at different times as input and outputs the flow map \bm{\mathcal{F}}_{ji} , occlusion map \bm{\mathcal{O}}_{ji} , and uncertainty map \bm{\mathcal{U}}_{ji} . Since the uncertainty of an estimation can also be interpreted as its variance, the initial flow prediction from frame j to frame i at location \bm{p} is computed as follows: \begin{split}(\bm{f}_{ji},\mu_{ji})=\begin{cases}(S(\bm{\mathcal{F}}_{ji},\bm{p}),S(\bm{\mathcal{U}}_{ji},\bm{p}))&\text{if}(\bm{\mathcal{O}}_{ji},\bm{p}))>\rho\\ None&\text{otherwise}\end{cases}\end{split} (1) where \rho=0.1 , and S(Target,\bm{p}) indicates sampling the target at location \bm{p} . The initial flow predictions are then set as input to flow integration. Following MFT, we adopt \{\infty,1,2,4,8,16,32\} as the time intervals, meaning that each frame’s prediction is computed by combining results from these previous frames. 2.3 Outlier Removal in Integration Even after dual filtering, occasional erroneous predictions may persist. To ensure a more stable integration process and minimize the impact of these incorrect predictions, we discard rough predictions that deviate significantly from others. When a rough prediction is more than 10 pixels away from the most trustworthy one, we mark it as wrong prediction. Denoting the input predictions as \{(\bm{f}_{1},\bm{\mu}_{1}),(\bm{f}_{2},\bm{\mu}_{2}),\dots,(\bm{f}_{N},\bm{\mu}_{N})\} , we remove outliers using the following criterion: (\bm{f}_{i},\bm{\mu}_{i})=\begin{cases}(\bm{f}_{i},\bm{\mu}_{i})&\text{if}\;\text{Norm}(\bm{f}_{i}-\bm{f}_{T})<\rho_{dist}\\ \text{None}&\text{otherwise}\end{cases} (2) where \rho_{dist}=10 , T=argmin(\mu_{i}) and \text{Norm}(\bm{x}) measures the magnitude of a vector. 3 Dense Inference As discussed in Sec.3.1 in the main paper, we utilize a geometry-aware feature extractor and a video mask generator for the dual-filter stage. While optical flow and geometry-aware features can be computed densely, generating masks for each pixel is both time-intensive and memory-intensive. To address this, we adopt an iterative approach to efficiently generate a set of masks that collectively cover all pixels, as described below: Algorithm 1 Dense Mask Generation Algorithm 1: Initialize: Set all pixels as unassigned. 2: while there exist unassigned pixels do 3: Select the first unassigned pixel p . 4: Generate a new mask M starting from pixel p . 5: for each pixel q in M do 6: if q is unassigned then 7: Assign q to mask M . 8: end if 9: end for 10: end while 11: Output: All pixels assigned to corresponding masks. Subsequently, the dual filter can be applied to each pixel based on its corresponding mask. 4 Training and Inference Speed Our methods is more than 20x faster than DINO-Tracker during the inference stage, while maintaining the same training time. The total time consumed for our method includes the time for keypoint extraction, mask generation, geometry-aware feature extraction and probabilistic integration. During keypoint extraction, we follow DINO-Tracker [ 49 ] to train a delta-DINO model and a heatmap refiner, which takes about 1 hour for an 80-frame video on a single RTX 4090 GPU. We refer to DINO-Tracker [ 49 ] for more details. However, our method skips the time-consuming occlusion prediction and directly uses points with high cosine similarity as keypoints, which saves much time. The mask generation and geometry-aware feature extraction together takes about 2 minutes and the probabilistic integration takes about 1 minute for the same video. In total, during the inference stage, the time spent tracking 3,000 points on a single object in an 80-frame video is about 3 minutes, which is about 20x faster than DINO-Tracker [ 49 ] . For dense inference, an additional 4 minutes may be required due to the increased number of masks generated, but our method remains more than 30x faster than DINO-Tracker [ 49 ] . Although our method is comprised of several components, we provide convenient interface to run our method directly. Our code will be released upon publication. 5 More Qualitative Results To further illustrate our methods’ robustness. We conduct experiments on more challenging cases and show the qualitative results. Some of the previous methods rely on computing a heatmap between the query point and the target frame. However, the per-frame heatmap lacks temporal-awareness and may confuse different objects. We address this issue by leveraging the mask and combining the heatmap with optical flow. As illustrated in Fig. 1 and Fig. 2 , by comparing the results of our method with DINO-Tracker [ 49 ] and TAPIR [ 13 ] , we show that although our method also relies on per-frame heatmap to extract keypoints,our method has strong temporal-awareness and is able to tell between similar objects. Figure 1 : Results of tracking a single object. While DINO-Tracker may mispredict parts onto similar objects and TAPIR can be disrupted by similar patterns, our method avoids these errors. Figure 2 : Results of tracking a single object. While DINO-Tracker may lose some parts and TAPIR can be disrupted by multiple similar patterns, our method avoids these errors. To further demonstrate the robustness of our method, we conduct experiments on extended videos from TAP-Vid-DAVIS, simulating high frame-rate videos by repeating each frame three times, as illustrated in Fig. 3 and Fig. 4 . In contrast to typical sliding-window or flow-based trackers (such as TAPTR [ 27 ] , SpatialTracker [ 52 ] and Co-Tracker [ 23 ] ), which tend to accumulate errors and drift over time, our integration of long-term key points with short-term optical flow enables continuous, drift-free tracking of the same point through occlusions. Figure 3 : Results of tracking at a higher frame rate. Sliding window based methods can easily lose track after occlusion and drift due to accumulating errors, while ours exhibit robustness. Figure 4 : Results of tracking at a higher frame rate. Sliding window based methods can mispredict points to other regions during occlusion (e.g. the gun and rope in shooting and the wrong person in india ), while ours exhibit robustness.