Track-On2 (fastforks)
Track-On2 / Track-On-R (Aydemir et al., 2025): online point tracking with a memory. A ViT-Adapter around a frozen DINOv3 ViT-S+/16 encodes each frame into 4 feature levels; each query decodes its position frame by frame against the frame features and its own memory of the last 72 frames, re-ranks the 16 best correlation peaks with deformable attention and refines the offset. Frames are processed causally (a streaming tracker). The DINOv3 weights are gated: accept the DINOv3 licence on Hugging Face, download model.safetensors yourself and put it under FASTFORKS_WEIGHTS_PATH (path and SHA-256 in the load error).
This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/gorkaydemir/track_on at 04efe8a137bd.
Author weight repos: https://huggingface.co/facebook/dinov3-vits16plus-pretrain-lvd1689m, https://huggingface.co/gorkaydemir/track_on2, https://huggingface.co/gorkaydemir/track_on_r.
Task: tracking. Tier: experimental. Outputs: tracks, visibility. Backends: portable, kernels.
Input: a video (its frames in order) and query points (t, x, y). Output: the position of every query in every frame and the probability that it is visible there.
Variants
Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.
| variant | weights (author link) | SHA-256 | weights licence | commercial use |
|---|---|---|---|---|
track-on-r@1 |
track_on_r.pt dinov3-vits16plus-pretrain-lvd1689m/model.safetensors |
a77165e6e84b208146e499da |
MIT AND LicenseRef-DINOv3-License | yes |
track-on2@1 |
trackon2_dinov3_checkpoint.pt dinov3-vits16plus-pretrain-lvd1689m/model.safetensors |
0e319c279cbd208146e499da |
MIT AND LicenseRef-DINOv3-License | yes |
Licence
Original-code strings, recorded as they are (research-use package; check them before any other use). Code: MIT AND LicenseRef-DINOv3-License. Licence files: LICENSE, LICENSE.md.
Training data: Upstream README / paper: Track-On2 trained on Kubric MOVi-F (synthetic, 64-frame clips); Track-On-R fine-tuned on real-world videos with pseudo-labels from an ensemble of trackers (verifier). Neither has a dataset entry here yet. The frozen DINOv3 ViT-S+/16 was pretrained on LVD-1689M (not released)..
- Licence note: The DINOv3 ViT-S+/16 weights and net/dinov3.py are under the DINOv3 License (Meta: attribution and conditions apply); read it before commercial use. The Track-On code and checkpoints are MIT.
Licence evidence: https://huggingface.co/gorkaydemir/track_on2/blob/672a47c4614f6d661edc36b7999a5bb5c3a5d2c9/README.md (license: mit); https://huggingface.co/gorkaydemir/track_on_r/blob/9401fe5c9ec6e21fb4f6070f05ac9ce75e3c9845/README.md (license: mit); https://github.com/gorkaydemir/track_on/blob/04efe8a137bd76e45e07382f02748794c9b61a96/LICENSE (MIT)
Input size
any H x W >= 32, padded to a multiple of 32
Tested up to (portable, fp16, RTX 5060 Ti 16 GB): track-on-r 3000x4000 (5.0 GiB peak allocated, 6.0 reserved); track-on2 3000x4000 (5.0 GiB peak allocated, 6.0 reserved)
Speed and memory
Tracking, synthetic, 480x854
| model | time (ms): original | portable | kernels | kernels speed-up: vs original | vs portable | peak memory (MiB, allocated / reserved): original | portable | kernels |
|---|---|---|---|---|---|---|---|---|
track-on-r |
2022 | 1188 | 1102 | 1.8x | 1.1x | 1801 / 1928 | 1423 / 1632 | 1368 / 1620 |
track-on2 |
1953 | 1194 | 1109 | 1.8x | 1.1x | 1403 / 1558 | 1423 / 1572 | 1368 / 1620 |
Startup and first call
| model | load (s): original | portable | kernels | first call (s): original | portable | kernels |
|---|---|---|---|---|---|---|
track-on-r |
0.88 | 0.24 | 1.11 | 2.91 | 1.99 | 1.72 |
track-on2 |
3.77 | 0.66 | 0.27 | 3.63 | 2.27 | 1.78 |
torch.compile and TensorRT
| model | input | time: portable | compile | TensorRT fp16 | kernels | preparation: compile (cold first call) | ONNX + engine build | peak MiB (alloc / reserved): portable | compile | kernels | TensorRT (engine + torch) |
|---|---|---|---|---|---|---|---|---|---|---|---|
track-on-r |
480x854 | 1188 ms | deferred (build >5 min) | planned for 0.2.1 | 1102 ms | β | β | 1423 / 1632 | β | 1368 / 1620 | β |
track-on2 |
480x854 | 1194 ms | deferred (build >5 min) | planned for 0.2.1 | 1109 ms | β | β | 1423 / 1572 | β | 1368 / 1620 | β |
Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).
Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.
Accuracy
On public datasets
TAP-Vid protocol: frames resized to 256 x 256 (PIL Lanczos) before the model (AllTracker: the original frames, resized by the model), one query per track at its first visible point, tracked online (TAPNext, Track-On) or over the whole clip (TAPIR, AllTracker, LocoTrack); AJ, delta_avg (< 1, 2, 4, 8, 16 px at 256 x 256) and occlusion accuracy averaged over the videos. Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | original fp16 | portable | kernels | note |
|---|---|---|---|---|---|---|
track-on2 |
TAP-Vid DAVIS (query first), every 3rd video, AJ / delta_avg / OA % | 69.60 / 81.26 / 92.96 | 69.64 / 81.30 / 92.94 | 69.69 / 81.33 / 92.94 | 69.65 / 81.32 / 92.93 |
Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks db9e79d5, 2026-11-13.
Parity with the original code
Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 1.8e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.05 and p99 <= 1.2 (absolute error of the output tracks).
Limitations
- Frames are resized to 384 x 512 (the size the learned position embedding was trained at), whatever their size and aspect ratio.
- Upstream adds a 20 x 20 grid of support points on frame 0 to every call (option support_grid); they are tracked with the user's queries, so a call costs at least 400 tracked points.
- Online: a query only sees the frames from its own frame on; frames before it return the query position with visibility 0.
- One video per call; all queries of a call share the video.
- Track-On2 peaks 1423 MB, 20 MB (1.4 %) above the fp16 original (1403 MB), and runs 15 % faster. The 1236 MB original variant is the fp32 one, which is 2.4x slower.
Usage
from fastforks import tracking
model = tracking.load("track-on-r") # mode="auto": the preferred backend from recorded measurements
result = model(frames, [[0, x, y]]) # frames: the video's images in order; queries (t, x, y)
result.tracks # [T, Q, 2] float32: x, y in input pixels; result.visible [T, Q] bool
result.backend # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')
Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: track-on-r, track-on2.
Kernels backend
tracking.load("track-on-r", mode="kernels") runs our CUDA kernels (fastbb, dpt_fused). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.