- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
PureTrack combines three ideas for efficient single-object tracking: purifying the initial template once, passing a compact spatio-temporal token between frames, and applying a bidirectional response-map consistency loss during training. The template is then cached rather than re-encoded at every frame. The paper reports 72.5 AO on GOT-10k and 208/76/84 FPS on GPU/CPU/Jetson AGX.
Critical strengths
-
Separating template initialization from online tracking is a sensible way to cut repeated computation. In Table 3, the decoupled version improves GPU speed from 162 to 211 FPS while also slightly improving AO.
-
The evaluation covers several tracking datasets and reports speed on three platforms. The AGX measurements are particularly relevant to the paper's efficiency objective.
Critical weaknesses
- The ablation study does not clearly separate the effects of the proposed components.
a) I had trouble reconciling Eq. (1) with Table 3. AGTP uses attention from the initial spatio-temporal token to the template tokens, yet configurations #3 and #7 include AGTP with the “s-t token” column unchecked. Does that column mean that cross-frame propagation is disabled, rather than that the token is absent? If so, the table should say so. This distinction matters when interpreting the small change from #7 (72.3 AO) to #8 (72.5 AO).
b) The evidence for the claimed complementarity between AGTP and BTC is also weak. AGTP improves AO from 71.2 to 71.8 when added to the decoupled model, but adds only 0.1 point when BTC is already present (72.2 to 72.3). The joint model is slightly better, but this alone does not establish a meaningful interaction between the two components.
c) Temporal token propagation is not new to tracking; ODTrack and LETrack are relevant examples. The paper cites related work, but I would like a clearer account of what the proposed token adds beyond these approaches, particularly given its 0.2-point gain in the reported ablation.
- The training setup and some reported comparisons need clarification.
a) Section 4.1 lists five training datasets, including LaSOT, TrackingNet, COCO2017, GOT-10k, and VASTTrack. It is not clear whether the GOT-10k result and the ablations in Table 3 come from a GOT-10k-only model or the jointly trained model. This should be stated explicitly because it affects how the results can be compared with other trackers.
b) The baseline in Table 3 already reaches 70.6 AO, but the source of the Fast-iTPN-Tiny pretrained weights and the baseline initialization are not specified clearly enough. Since the full method improves on this baseline by 1.9 points, these details are important for interpreting the gains and reproducing the results.
c) There is also a concrete error in the TrackingNet discussion on page 6. ARTrack is called the fastest non-real-time tracker, and PureTrack is said to be 4.4 times faster than ARTrack on AGX. Table 1 instead gives 84 FPS for PureTrack and 8 FPS for ARTrack-256, a ratio of 10.5. The 4.4 ratio matches the comparison with OSTrack-256 (19 FPS), which is itself faster than ARTrack-256 on all three reported platforms. The text appears to mix up the two trackers and should be corrected.
Questions for authors
-
In Table 3, does “s-t token” refer specifically to recurrent propagation? What token is used to compute AGTP in configurations #3 and #7?
-
What is the main technical difference between PureTrack's temporal token and those used in ODTrack and LETrack? Is there additional evidence of a benefit beyond the 0.2 AO gain in Table 3?
-
Were the GOT-10k results and ablations obtained with GOT-10k-only training? What pretrained weights and initialization were used for the 70.6 AO baseline?
-
Please check the ARTrack/OSTrack speed comparison on page 6 against Table 1 and correct the ranking and speed ratio.
-
Some Mistakes a) The Figure 3 caption on page 9 begins “isualization”; it should read “Visualization”. The abbreviations “w” and “w/o” could also be made consistent. b) The conclusion contains “spatio-temporal token serve”; the number and verb agreement should be corrected.