Models · Point tracking

TAPIR

TAPIR and BootsTAPIR (Google DeepMind, 2023 / 2024): track any point through a whole video offline: a per-frame match from a cost volume, then four temporal refinement passes; BootsTAPIR adds self-training on real videos. Apache-2.0 code and weights.

experimental Fork of google-deepmind/tapnet@ceedd60 · backends: portable, kernels

Checkpoint

Paper 2024 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 192 ms 480x854, fp16
vs the faster original 1.3× original fp16: 243 ms
Peak memory 508 MiB original: 639 MiB
Accuracy vs original fp32 +0.02 kernels aj, TAP-Vid DAVIS (query first) (65.49 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import tracking

model = tracking.load("bootstapir")  # 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 bootstapir

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)

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32not measured ––––
Original fp16the authors' mixed-precision option (float16); upstream BootsTAPIR, offline (whole video); frames resized to 256 x 256fp16 243–639 MiB/774 MiB0.50.46
Portable fp32fp32not measured ––––
Portablefp16 2341.0×508 MiB/628 MiB0.30.80
Kernelsfp16 1921.3×508 MiB/628 MiB1.10.66
Portable + torch.compile 1831.3×758 MiB/832 MiB207 (build)–
TensorRT fp16fp16 13.218×204 MiBenginenot comparable65 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-12 to 2026-11-13. 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

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp321930 2,931 MiB process–480x854
Portablefp162059 1,750 MiB processmatches480x854
Portablefp321721 2,769 MiB processmatches480x854
Kernelsfp16not 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
Apache-2.0 LICENSE
Weights
Apache-2.0 source
Training data
Upstream: TAPIR is trained on synthetic Kubric MOVi-E sequences (the panning variant); BootsTAPIR is further trained with BootsTAP self-training on about 15 million unlabelled real video clips (not released). None has a dataset entry here yet.

Our reading of the licences, not legal advice.

Weights

bootstapir_checkpoint_v2.pt 219 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download bootstapir).

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 %

Rundtypeaj ↑delta_avg ↑oa ↑Δ aj vs fp32 originalvideosms / video
Original fp32fp3265.4976.7592.56 reference 10 –
Original fp16fp1665.5376.7292.55 +0.04 10 –
Portable fp32fp3265.4976.7692.58 ±0.00 10 –
Portablefp1665.5676.7892.44 +0.07 10 –
Kernelsfp1665.5176.7292.51 +0.02 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, 2026-11-12; 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 8d5bccf, 2026-11-12.

TAP-Vid DAVIS (query first)

AJ / delta_avg / OA %

Rundtypeaj ↑delta_avg ↑oa ↑Δ aj vs fp32 originalvideosms / video
Original fp32fp3262.2774.6289.67 reference 30 –
Original fp16not measured
Portable fp32fp3262.2674.5989.68 −0.01 30 –
Portablefp1662.2974.6189.66 +0.02 30 –
Kernelsfp1662.2774.6189.64 ±0.00 30 –

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: 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 8d5bccf, 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| <= 3.5e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0079 and p99 <= 0.028 (absolute error of the output tracks).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000035 not recorded
GPU default dtypes, mean |diff|≤ 0.0079 0.00524
GPU default dtypes, p99 |diff|≤ 0.028 0.01845

Recorded for bootstapir on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12, with these kernels: fvm_conv

Details

TAPIR (Doersch, Yang, Vecerik et al., Google DeepMind 2023) and BootsTAPIR (Doersch et al. 2024, the same network trained further with BootsTAP self-training on unlabelled real videos): tracking any point offline over a whole video. A ResNet encodes every 256 x 256 frame; per frame, a cost volume against the query's features gives the initial position, visibility and uncertainty, then four PIPs-style refinements (local correlations, a depthwise-conv MLP mixer over the frames of each track) correct the whole track at once.

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 (5,747 MiB allocated)
Outputs
tracks, visibility
CUDA kernels
  • fvm_convfused convolutions for image classifiers
Apple (MPS)
fp32, fp16

Limitations

  • Offline: the whole video is encoded before tracking (features of every frame in memory, about 3 MB per frame); for live video use an online tracker (TAPNext).
  • Frames are resized to 256 x 256 (the training size and that of upstream's TAP-Vid numbers), whatever their size and aspect ratio. Upstream's higher-resolution refinement (best at 512 x 512 for BootsTAPIR) is not ported.
  • One video per call; all queries of a call share the video.
Full reference, generated from the repository

TAPIR and BootsTAPIR (Google DeepMind, 2023 / 2024): track any point through a whole video offline: a per-frame match from a cost volume, then four temporal refinement passes; BootsTAPIR adds self-training on real videos. Apache-2.0 code and weights.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
bootstapirsynthetic (480x854)original
portable
kernels
portable + torch.compile
TensorRT fp16
243
234
192
183
13.2
639 / 774
508 / 628
508 / 628
758 / 832
492 / 532
tapirsynthetic (480x854)original
portable
kernels
portable + torch.compile
TensorRT fp16
187
182
141
131
6.47
498 / 574
463 / 548
463 / 548
630 / 752
361 / 406
Details
  • Original code: https://github.com/google-deepmind/tapnet at ceedd60bb9f6
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
  • Licence evidence: https://github.com/google-deepmind/tapnet/blob/ceedd60bb9f6c9f7c74796e4445aff6c6ddf7c9b/README.md#license-and-disclaimer
  • Training data: Upstream: TAPIR is trained on synthetic Kubric MOVi-E sequences (the panning variant); BootsTAPIR is further trained with BootsTAP self-training on about 15 million unlabelled real video clips (not released). None has a dataset entry here yet.
  • Kernel libraries: fvm_conv
  • Variants: bootstapir, tapir
  • 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): bootstapir 3000x4000 (5.6 GiB peak allocated, 6.1 reserved); tapir 3000x4000 (5.5 GiB peak allocated, 6.0 reserved)
  • bootstapir 480x854, first call: portable + torch.compile 207 s (cold Inductor cache), TensorRT fp16 65 s (ONNX + engine).
  • tapir 480x854, first call: portable + torch.compile 169 s (cold Inductor cache), TensorRT fp16 55 s (ONNX + engine).
  • bootstapir original 480x854: the authors' mixed-precision option (float16); upstream BootsTAPIR, offline (whole video); frames resized to 256 x 256
  • tapir original 480x854: the authors' mixed-precision option (float16); upstream TAPIR, offline (whole video); frames resized to 256 x 256
  • bootstapir load / first call: original 0.6 s / 0.901 s, portable 0.32 s / 0.8 s, kernels 1.14 s / 0.659 s
  • tapir load / first call: original 0.35 s / 0.827 s, portable 0.19 s / 0.836 s, kernels 1.13 s / 0.688 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.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
bootstapirTAP-Vid DAVIS (query first), every 3rd video, AJ / delta_avg / OA %65.49 / 76.75 / 92.5665.53 / 76.72 / 92.5565.56 / 76.78 / 92.4465.49 / 76.76 / 92.5865.51 / 76.72 / 92.51
bootstapirTAP-Vid DAVIS (query first), AJ / delta_avg / OA %62.27 / 74.62 / 89.6762.29 / 74.61 / 89.6662.26 / 74.59 / 89.6862.27 / 74.61 / 89.64
tapirTAP-Vid DAVIS (query first), every 3rd video, AJ / delta_avg / OA %62.63 / 74.54 / 90.1462.62 / 74.60 / 90.1862.74 / 74.59 / 90.1162.62 / 74.51 / 90.1462.73 / 74.55 / 90.21

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 8d5bccf, 2026-11-12 to 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 8d5bccf, 2026-11-12 to 2026-11-13.

Parity

CPU, float32: max |diff| <= 3.5e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0079 and p99 <= 0.028 (absolute error of the output tracks).

variantkernels error vs original, mean / p99GPU, torchkernels used
bootstapir0.00524 / 0.01845RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120
tapir0.0036 / 0.01425RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120

Limitations

  • Offline: the whole video is encoded before tracking (features of every frame in memory, about 3 MB per frame); for live video use an online tracker (TAPNext).
  • Frames are resized to 256 x 256 (the training size and that of upstream's TAP-Vid numbers), whatever their size and aspect ratio. Upstream's higher-resolution refinement (best at 512 x 512 for BootsTAPIR) is not ported.
  • One video per call; all queries of a call share the video.

Example output

No example rendered for this checkpoint yet.