Track-On2
Track-On2 and Track-On-R (Aydemir et al., 2025): online point tracking, one frame at a time, as a transformer decoder over a DINOv3 ViT-S+ adapter with a memory of past features; -R is fine-tuned on real videos. MIT; DINOv3 weights under the DINOv3 License.
experimental Fork of gorkaydemir/track_on@04efe8a · backends: portable, kernels
- Paper Track-On2: Enhancing Online Point Tracking with Memory TPAMI 2026 · first released 23 Sept 2025
Paper 2025 Added in 0.2.0 Evaluated with 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install fastforks-core
from fastforks import tracking
model = tracking.load("track-on2") # kernels on a supported GPU, else portable PyTorch
result = model(video, queries)
result.tracks # also: visibility
result.backend # what ran: mode, device, dtype, kernels fastforks bench track-on2 Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.
Speed and memory
Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.
Input: synthetic, 480x854 (the task’s standard size)
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | not measured | – | – | – | – |
| Original fp16the authors' mixed-precision option (float16); gorkaydemir/track_on@04efe8a137bd Predictor loop, 384 x 512, memory 72 | fp16 | 1953 | – | 1,403 MiB/1,558 MiB | 0.5 | 3.07 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 1193 | 1.6× | 1,423 MiB/1,712 MiB | 1.2 | 1.96 |
| Kernels | fp16 | 1108 | 1.8× | 1,368 MiB/1,620 MiB | 0.6 | 1.76 |
| Portable + torch.compiletimeout after 900 s (audit 2026-11-16, RTX 5060 Ti: compile | fails | – | – | – | – | |
| TensorRTplanned for 0.2.1 | fails | – | – | – | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-13 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two
times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.
Other devices
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-13 | Original | fp32 | 11323 | 3,052 MiB process | – | 480x854 |
| Portable | fp32 | 8354 | 6,102 MiB process | matches | 480x854 | |
| Kernels | fp16 | not implemented | – | 480x854 | ||
CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.
Licence and weights
Commercial use allowed
- Code
- MIT AND LicenseRef-DINOv3-License LICENSELICENSE.md
- Weights
- MIT AND LicenseRef-DINOv3-License source
- 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).
- 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.
Our reading of the licences, not legal advice.
Accuracy
The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.
TAP-Vid DAVIS (query first)
every 3rd video, AJ / delta_avg / OA %
| Run | dtype | aj ↑ | delta_avg ↑ | oa ↑ | Δ aj vs fp32 original | videos | ms / video |
|---|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 69.60 | 81.26 | 92.96 | reference | 10 | – |
| Original fp16 | fp16 | 69.64 | 81.30 | 92.94 | +0.04 | 10 | – |
| Portable fp32 | fp32 | 69.59 | 81.26 | 92.95 | −0.01 | 10 | – |
| Portable | fp16 | 69.69 | 81.33 | 92.94 | +0.09 | 10 | – |
| Kernels | fp16 | 69.65 | 81.32 | 92.93 | +0.05 | 10 | – |
Query-first 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 authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks bb4e8b64, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks db9e79d5, 2026-11-13.
Closeness to the original
Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. 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).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.000018 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.05 | 0.02559 |
| GPU default dtypes, p99 |diff| | ≤ 1.2 | 0.49624 |
Recorded for track-on2 on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: dpt_fusedfastbb
Details
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).
- Input / output
- 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. - Input size
- any H x W >= 32, padded to a multiple of 32; tested up to 3000x4000 (5,124 MiB allocated)
- Options
memory=72>= 1· changes the outputsupport_grid=20>= 0· changes the output
- Outputs
tracks,visibility- CUDA kernels
fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbonesdpt_fusedfused fp16 matrix multiplies and convolutions for DINOv2 backbones and DPT depth heads
- Apple (MPS)
- fp32
Limitations
- At 480 x 854 portable / kernels peak 1423 / 1368 MB, above the float32 original (1236 MB) but 2.4x / 2.6x faster (1194 / 1109 vs 2861 ms); the original under fp16 autocast needs 1403 MB, like portable.
- 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.
- No exportable module: every step reads a data-dependent scalar (`torch.compile` runs with graph breaks; ONNX / TensorRT export fails).
Full reference, generated from the repository
Track-On2 and Track-On-R (Aydemir et al., 2025): online point tracking, one frame at a time, as a transformer decoder over a DINOv3 ViT-S+ adapter with a memory of past features; -R is fine-tuned on real videos. MIT; DINOv3 weights under the DINOv3 License.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
track-on2 | synthetic (480x854) | original portable kernels portable + torch.compile TensorRT fp16 | 1953 1193 1108 fails: timeout after 900 s (audit 2026-11-16, RTX 5060 Ti: compile planned for 0.2.1 | 1403 / 1558 1423 / 1712 1368 / 1620 – – |
track-on-r | synthetic (480x854) | original portable kernels portable + torch.compile TensorRT fp16 | 2022 1188 1102 fails: timeout after 900 s (audit 2026-11-16, RTX 5060 Ti: compile planned for 0.2.1 | 1801 / 1928 1423 / 1632 1368 / 1620 – – |
Details
- Original code: https://github.com/gorkaydemir/track_on at
04efe8a137bd - Code licence: MIT AND LicenseRef-DINOv3-License (permissive); files:
LICENSE,LICENSE.md - Weights licence: MIT AND LicenseRef-DINOv3-License
- 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)
- 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.
- Options:
memory= 72 (>= 1),support_grid= 20 (>= 0) - Kernel libraries:
fastbb,dpt_fused - Variants:
track-on2,track-on-r - Not in this release (planned for 0.2.1):
track-on2ONNX / TensorRT export: torch.export stops at a data-dependent branch (model.py:107);track-on-rONNX / TensorRT export: the export fails in the network call (ValueError: expected 4 values, got 2) - Outputs: tracks, visibility
- Input size: any H x W >= 32, padded to a multiple of 32
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
track-on23000x4000 (5.0 GiB peak allocated, 6.0 reserved);track-on-r3000x4000 (5.0 GiB peak allocated, 6.0 reserved) track-on2original 480x854: the authors' mixed-precision option (float16); gorkaydemir/track_on@04efe8a137bd Predictor loop, 384 x 512, memory 72track-on-roriginal 480x854: the authors' mixed-precision option (float16); gorkaydemir/track_on@04efe8a137bd Predictor loop, 384 x 512, memory 72track-on2load / first call: original 3.77 s / 3.626 s, portable 1.16 s / 1.956 s, kernels 0.63 s / 1.762 strack-on-rload / first call: original 0.88 s / 2.907 s, portable 0.24 s / 1.989 s, kernels 1.11 s / 1.72 s- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Measured
Query-first 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 | portable fp32 | 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 | 69.59 / 81.26 / 92.95 | 69.65 / 81.32 / 92.93 |
Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks db9e79d5, 2026-11-13; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks bb4e8b64, 2026-11-17; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks db9e79d5, 2026-11-13.
Parity
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).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
track-on2 | 0.02559 / 0.49624 | RTX 5060 Ti, torch 2.14.1+cu130 | dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120 |
track-on-r | 0.03312 / 0.75115 | RTX 5060 Ti, torch 2.14.1+cu130 | dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120 |
Limitations
- At 480 x 854 portable / kernels peak 1423 / 1368 MB, above the float32 original (1236 MB) but 2.4x / 2.6x faster (1194 / 1109 vs 2861 ms); the original under fp16 autocast needs 1403 MB, like portable.
- 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.
- No exportable module: every step reads a data-dependent scalar (
torch.compileruns with graph breaks; ONNX / TensorRT export fails).
Example output
No example rendered for this checkpoint yet.