--- kind: concept slug: dense-correspondence title: Dense correspondence summary: >- Matching pixels between images, or points on 3D objects and bodies. Covers label-free diffusion features (DIFT), one Transformer matcher for flow, stereo and depth (UniMatch), coarse-to-fine matching on frozen features (RoMa), GAN-feature tracking (DragGAN), descriptor fields for pose search (NDF) and body-object correspondence (BEHAVE). Trade-off: label-free convenience against per-task tuning. tags: [correspondence, diffusion-features, feature-extraction, semantic-correspondence, optical-flow, stereo-matching, dense-matching, point-tracking, descriptor-fields, human-object-interaction] papers: ["2306.03881", "2211.05783", "2305.15404", "2305.10973", "2112.05124", "2204.06950"] related: [diffusion-models, latent-diffusion, self-supervised-learning, local-feature-matching, optical-flow, cost-volume, gaussian-process, transfer-learning, shifted-window-attention, multi-view-stereo, latent-space-editing, euclidean-equivariant-network, implicit-neural-representation, neural-implicit-surface, multi-view-reconstruction, learning-from-demonstration] status: draft updated_at: 2026-11-01 --- ## Overview Dense correspondence gives, for a pixel in a source image, the pixel in a target image that shows the same point. Reconstruction, tracking, segmentation and editing all need it [arxiv:2306.03881 §1]. The usual route is a network trained on annotated pairs, which are costly to collect [arxiv:2306.03881 §2]. Optical flow, rectified stereo and posed-image depth ask the same matching question, and one model can answer all three [arxiv:2211.05783 §1]. Inside an editing tool, the same question appears as tracking: a handle the user placed must be found again after each change to the image [arxiv:2305.10973 §3.3]. Points on 3D surfaces raise the same question. NDF matches query points of a rigid body to a demonstrated pose through learned descriptors [arxiv:2112.05124 §II-B]. BEHAVE predicts, for points near a person and an object, their correspondence on a standard body model, and fits both models to them [arxiv:2204.06950 §4.1]. Here "correspondence" means a point on a model surface, not a pixel in another image. The image matchers share one skeleton: compute a feature for every pixel, score every candidate in the other image, then turn the scores into a match. The variants differ in where the features come from and how the scores become a match. DIFT takes the nearest neighbour of features read from a pretrained diffusion denoiser [arxiv:2306.03881 §3]. UniMatch takes a softmax-weighted average over all target pixels, with no learned matching parameters [arxiv:2211.05783 §3.1]. RoMa predicts a distribution over a 64 × 64 anchor grid for each coarse pixel, then refines the warp [arxiv:2305.15404 §3.3]. DragGAN takes the nearest neighbour, but only inside a small square around the handle, using features of a StyleGAN2 generator [arxiv:2305.10973 §3.3]. The baselines in the same tables are self-supervised image encoders (DINO, OpenCLIP), used without fine-tuning, and supervised or weakly supervised matchers [arxiv:2306.03881 §5.1] [tab. 1]. They are comparison points here, not variants. ## Variants ### Diffusion features on a Stable Diffusion backbone (DIFT_sd) A real image is noised to level t and passed once through the denoiser. The output of one upsampling block is kept as the feature map. The backbone is Stable Diffusion 2-1, trained on LAION [arxiv:2306.03881 §5.1]. The paper describes noising the image [arxiv:2306.03881 §4.2]; the code encodes it with the VAE and noises the latent instead [code: src/models/dift_sd.py#L177-L181]. On SPair-71k this is the strongest row without labels: 52.9 PCK per image and 59.5 per point [arxiv:2306.03881 tab. 1] [tab. 2]. Its settings for the semantic tasks are t = 261 and block 1, at 768 × 768 [arxiv:2306.03881 app. C]. Card: [Emergent Correspondence from Image Diffusion](/papers/2306.03881). ### Diffusion features on an ADM backbone (DIFT_adm) The same procedure with the Ablated Diffusion Model, trained on ImageNet without labels [arxiv:2306.03881 §5.1]. Its settings for the semantic tasks are t = 101 and block 4, at 512 × 512 [arxiv:2306.03881 app. C]. It is the stronger option on video and is weaker on SPair-71k: 46.3 PCK per image and 52.0 per point [arxiv:2306.03881 tab. 1] [tab. 2]. | Task | Backbone | Time step t | Block n | Input | Source | |---|---|---|---|---|---| | SPair-71k, PF-WILLOW, CUB | Stable Diffusion | 261 | 1 | 768 × 768 | [arxiv:2306.03881 app. C] | | SPair-71k, PF-WILLOW, CUB | ADM | 101 | 4 | 512 × 512 | [arxiv:2306.03881 app. C] | | HPatches | Stable Diffusion | 0 | 2 | 768 × 768 | [arxiv:2306.03881 app. C] | | HPatches | ADM | 41 | 11 | 768 × 768 | [arxiv:2306.03881 app. C] | The paper chooses t and n by grid search per task [arxiv:2306.03881 §4.2] [app. C]. For SPair-71k the search is on that set, and the same values are reused for PF-WILLOW and CUB [arxiv:2306.03881 §5.2]. ### Label propagation for video For video, a label propagation scheme passes labels from earlier frames to later ones. The propagation uses a softmax temperature, a top-k filter and a spatial radius, with a set number of previous frames [arxiv:2306.03881 §6.2]. The DAVIS settings are 28 previous frames and a radius of 15 pixels [arxiv:2306.03881 app. C]. The paper reports that features alone do not reproduce the video numbers; the propagation is part of the method [arxiv:2306.03881 §6.2]. ### Transformer matcher for flow, stereo and depth (UniMatch) A weight-shared CNN gives 1/8-resolution features for each image. Six Transformer blocks, each with self-attention, cross-attention and a feed-forward layer, then enhance both feature maps, with fixed sine and cosine positions [arxiv:2211.05783 §3.2]. Attention runs in local windows, shifted by half a window, not over all pixels [arxiv:2211.05783 tab. 2]. The matching layer has no parameters. Every pixel is correlated with the other view, a softmax over all targets gives a distribution, and its expected coordinate gives flow, disparity or depth [arxiv:2211.05783 §3.1] [eq. 1–8]. The feature network and the propagation layer are shared across tasks; only this matching step is task-specific [arxiv:2211.05783 §3.1]. Pixels with no clear match, such as occluded ones, take values from a self-attention layer over the features [arxiv:2211.05783 §3.3]. Optional refinements add a 1/4-resolution matching step and RAFT-style local iterations. The final models use 6 local steps for flow, 3 for stereo and 1 at 1/8 for depth [arxiv:2211.05783 §3.4.2] [§4.5]. On Sintel, GMFlow with one refinement beats RAFT with 31 refinements on training EPE: 1.08 against 1.41 on clean and 2.48 against 2.69 on final. The times are 151 ms against 170 ms on a V100 at 436 × 1024 [arxiv:2211.05783 tab. 3]. For stereo, 1D horizontal cross-attention is faster and more accurate than 2D: Things EPE 1.22 at 50 ms against 1.25 at 61 ms [arxiv:2211.05783 §3.2] [tab. 6]. Starting from a flow model helps the other tasks. For stereo on Things, EPE with no fine-tuning is 2.58 from flow and 75.24 from random initialisation; after 100K steps it is 1.00 against 1.11. For depth on ScanNet, Abs Rel with no fine-tuning is 0.198 against 0.536, and 0.063 against 0.069 after 100K steps [arxiv:2211.05783 tab. 10]. The paper finds no gain in the other direction [arxiv:2211.05783 §4.4]. Card: [Unifying Flow, Stereo and Depth Estimation](/papers/2211.05783). ### Frozen foundation features with a coarse-to-fine matcher (RoMa) RoMa is a coarse-to-fine matcher built on DKM's structure [arxiv:2305.15404 §3.1]. Its coarse features are the patch tokens of a frozen DINOv2 ViT-L/14, projected to 512 channels [arxiv:2305.15404 §3.2] [app. B]. Its fine features come from a VGG19 pyramid at strides 1, 2, 4 and 8 [arxiv:2305.15404 app. B]. The authors chose VGG19 for this role after an ablation [arxiv:2305.15404 §3.2] [tab. 2]. A Gaussian-process match encoder and a 5-block Transformer decoder with no positional encodings then predict, for each coarse pixel, logits over 64 × 64 anchors plus a matchability logit [arxiv:2305.15404 §3.3] [app. B]. Five convolutional refiners at strides 1, 2, 4, 8 and 14 move the warp from coarse to fine, with gradients detached between them [arxiv:2305.15404 §3.1] [app. B]. The coarse stage is trained as classification on the anchor nearest the target, plus a matchability term. The fine stage uses a generalised Charbonnier loss with α = 0.5. The two are summed with no extra weights [arxiv:2305.15404 §3.4]. Training needs dense supervised correspondences from MVS depth (MegaDepth) or RGB-D (ScanNet) [arxiv:2305.15404 §4.2]. Author-reported results. On WxBS, mAA at 10 px is 80.1, against 58.9 for DKM and 55.4 for LoFTR [arxiv:2305.15404 tab. 4]. On the IMC2022 hidden test set, mAA at 10 is 88.0, against 83.1 for DKM and 83.8 for ASpanFormer [arxiv:2305.15404 tab. 3]. On MegaDepth-1500, AUC at 5° is 62.6, against 60.4 for DKM [arxiv:2305.15404 tab. 5]. In the ablation, the DKM row is retrained by the authors, not the published model [arxiv:2305.15404 §4.1] [tab. 2]. The per-pair runtime at 560 × 560 is 198.8 ms against 186.3 ms for DKM, on an RTX6000 [arxiv:2305.15404 §4.5]. Card: [RoMa: Robust Dense Feature Matching](/papers/2305.15404). ### Nearest-neighbour tracking in GAN features (DragGAN) DragGAN edits a StyleGAN2 image by moving user-chosen handle points to target points. Each iteration takes one optimiser step on the latent code. Then each handle moves to the pixel inside a square around it whose block-6 feature is nearest, in L1 distance, to the feature the handle had at the start [arxiv:2305.10973 §3.3] [eq. 2]. The square's half-width is 12 pixels at 512 px [arxiv:2305.10973 §3.4]. No tracker network is used, and the same features also drive the motion [arxiv:2305.10973 §3.2]. Swapping the tracker changes the result. With RAFT tracking, the one-point face-landmark distance is 13.43, and with PIPs it is 2.98, against 2.44 for the full method [arxiv:2305.10973 tab. 1]. The authors report that the other trackers drift off the handle [arxiv:2305.10973 §4.1] [fig. 6]. These numbers measure the whole edit on face landmarks, not the tracker alone [arxiv:2305.10973 §4.2]. Card: [Drag Your GAN](/papers/2305.10973). ### Point descriptor fields for pose search (NDF) NDF gives each 3D point a descriptor. The descriptor is the concatenated activations of an occupancy network, so the network is trained by 3D reconstruction and needs no keypoint labels [arxiv:2112.05124 §II-A] [eq. 4]. Vector Neurons make the descriptor SE(3)-equivariant: rotating and moving the object moves the descriptors with it [arxiv:2112.05124 §II-A] [eq. 5–7]. A rigid body is represented by a set of query points, and its pose descriptor is the set of their descriptors [arxiv:2112.05124 §II-B] [eq. 10]. To find a pose, an optimiser starts from random poses and minimises the L1 distance to the demonstration's descriptor [arxiv:2112.05124 §II-B] [eq. 11]. The paper trains one model across mugs, bowls and bottles on 100,000 simulated objects [arxiv:2112.05124 §III]. Author-reported simulation results: with ten demonstrations, overall success on unseen upright objects is 0.88 for mugs, 0.91 for bowls and 0.87 for bottles, and 0.58 to 0.78 for arbitrary poses. The Dense Object Nets baseline reaches 0.00 to 0.45 on the same tasks [arxiv:2112.05124 tab. I]. With one demonstration on upright mugs, overall success is 0.46, against 0.32 for Dense Object Nets [arxiv:2112.05124 tab. IV]. Real-robot results are qualitative only [arxiv:2112.05124 §IV-C]. Card: [Neural Descriptor Fields](/papers/2112.05124). ### Body and object correspondence from multi-view depth (BEHAVE) BEHAVE tracks a person and a rigid object with a known template mesh, frame by frame, from four calibrated RGBD cameras [arxiv:2204.06950 §3] [§4]. A voxel-based network predicts, at each sampled 3D point, the unsigned distance to the human and to the object, a correspondence onto the SMPL body model, and the object orientation [arxiv:2204.06950 §4.1] [§4.2]. The fit combines a body term, an object term and a contact term that pulls touching object vertices onto their predicted body points [arxiv:2204.06950 eq. 1] [eq. 5]. Author-reported: SMPL vertex error is 4.99 cm and object vertex error 21.20 cm. The single-image PHOSA baseline gets 13.73 cm and 34.73 cm [arxiv:2204.06950 tab. 2]. Removing the contact prediction leaves SMPL error at 4.96 cm and object error at 21.28 cm, so the authors describe the contact term's effect as qualitative [arxiv:2204.06950 tab. 3] [§5.4]. The reference fits are pseudo-ground-truth, so these errors are relative to them [arxiv:2204.06950 §3]. Card: [BEHAVE](/papers/2204.06950). ## When to use which | Situation | Use | Why | |---|---|---| | No correspondence labels, and pairs may differ in category | DIFT_sd | Highest unsupervised row on SPair-71k: 52.9 per image against 38.4 for OpenCLIP and 33.9 for DINO [arxiv:2306.03881 tab. 1] | | Labelled pairs exist and the best accuracy on SPair-71k matters | A supervised method | SCorrSAN gets 55.3 per image, 2.4 points above DIFT_sd [arxiv:2306.03881 tab. 1] | | Homography on HPatches, with no geometric supervision | DIFT_sd at ε = 1; a weakly supervised keypoint method at ε = 3 or 5 | DIFT_sd gets 45.6 at ε = 1, above CAPS at 44.8. CAPS, with SuperPoint keypoints, gets 76.3 and 85.2 at ε = 3 and 5, against 73.9 and 83.1 for DIFT_sd [arxiv:2306.03881 tab. 4] | | Video, no video training data | DIFT_adm with label propagation | 75.7 DAVIS J&F, the highest row in the paper's table, above SFC at 71.2, which is trained on video [arxiv:2306.03881 tab. 5] | | GPU memory is short | Four noise copies, not eight | The paper's fallback for lower memory [arxiv:2306.03881 app. C] | | Weights must be licensed for commercial use | Check both checkpoints first | The paper's CC BY 4.0 covers the paper, not the Stable Diffusion or ADM weights, whose terms are not checked here [arxiv:2306.03881 app. B] | | Two views with hard appearance change, and labelled 3D scenes exist | RoMa | Above DKM on WxBS and IMC2022 in the paper's tables, the two hard wide-baseline sets it reports [arxiv:2305.15404 tab. 3] [tab. 4] | | Retraining the matcher from the paper | Plan for gaps | The paper says only that the setup follows DKM, and the loss weight λ is not stated [arxiv:2305.15404 §3.4] [§4.2] | | One model for flow, stereo and depth, or a flow start for stereo or depth | UniMatch | Flow-initialised stereo reaches 1.00 Things EPE after 100K steps, against 1.11 from random [arxiv:2211.05783 tab. 10] | | Optical flow with one refinement | UniMatch with one refinement | Sintel clean 1.08 against 1.41 for RAFT with 31 refinements, trained on Chairs and Things [arxiv:2211.05783 tab. 3] | | Flow on KITTI-like data, with only Chairs and Things for training | Expect UniMatch to trail RAFT | EPE 7.77 against 5.32 for RAFT; adding Virtual KITTI 2 narrows it to 2.85 against 2.45 [arxiv:2211.05783 tab. 5] | | Occluded or unmatched pixels matter | UniMatch's propagation layer | Sintel clean EPE on unmatched pixels falls from 15.54 to 10.39 [arxiv:2211.05783 tab. 2] | | Drag a point of a generated image to a target, with no annotated data and no tracker network, and a StyleGAN2 generator at hand | DragGAN-style tracking in GAN features | One-point face-landmark distance 2.44, against 11.64 for UserControllableLT, at 2.0 s per edit [arxiv:2305.10973 tab. 1]. The released code has a non-commercial research licence, so for a commercial product implement from the paper | | Grasp or place pose for a new object instance from a few demonstrations, with depth point clouds and no keypoint labels | NDF | Overall simulated success 0.88 for upright mugs with ten demonstrations, against 0.45 for Dense Object Nets [arxiv:2112.05124 tab. I]. Training needs 100,000 simulated objects [arxiv:2112.05124 §III], and real-robot evidence is qualitative only [§IV-C] | | Tracking a person and a known rigid object together, from calibrated multi-view RGBD | BEHAVE's per-frame fit | SMPL error 4.99 cm and object error 21.20 cm, against 13.73 cm and 34.73 cm for the single-image PHOSA [arxiv:2204.06950 tab. 2]. No temporal model is described [§4], and the reference fits are pseudo-ground-truth [§3]. Its code and dataset are non-commercial, so implement from the paper | | Real-time, 20 FPS or more | None of the image matchers | The UniMatch authors say the full model is not real-time [arxiv:2211.05783 §5]. RoMa takes about 199 ms per pair in the paper's setup [arxiv:2305.15404 §4.5] | Timings are in the paper's appendix: DIFT_sd takes 203 ms per 768 × 768 image on one A6000 and DIFT_adm 110 ms per 512 × 512 image. The DINO and OpenCLIP times are at other sizes, so treat the comparison as the paper's setup only [arxiv:2306.03881 app. B]. ## Implementation sketch The DIFT core is one denoiser pass per noise copy, then nearest-neighbour matching. The UniMatch core is a softmax over all target pixels, with no learned parameters, so both can be written in a few lines. The DragGAN tracker is the DIFT matcher restricted to a square window around each handle, with an L1 distance [arxiv:2305.10973 §3.3] [eq. 2]. Its pseudocode is on the card. ```pseudocode # Extract one feature map from a real image [arxiv:2306.03881 §4.2] [app. C] # Input: image in [-1, 1]; time step t; block index n; copies K # Output: feature map (C, h, w) def extract_features(image, t, n, copies): # Each copy gets its own noise draw, so repeat the image [app. C] batch = repeat(image, copies) # Noise every copy to level t with the forward process [§4.2] noisy = add_noise(batch, t) # One denoiser pass; keep the output of upsampling block n [app. C] acts = unet_up_blocks(noisy, t)[n] # Average over copies to get one feature map [§4.2] return mean(acts, axis=0) ``` ```pseudocode # Match each source point to the target image [§3] [eq. 1] # Input: source and target feature maps; source points # Output: matched target point for each source point def match(src_feat, tgt_feat, points): # Read the source feature at each point, bilinear interpolation [§3] queries = sample_bilinear(src_feat, points) # Cosine similarity with every target location [§3] scores = cosine(tgt_feat, queries) # Nearest neighbour is the highest-scoring target location [eq. 1] return argmax(scores, over=target_locations) ``` ```pseudocode # Soft matching for flow, stereo or depth [arxiv:2211.05783 §3.1] [eq. 1-4] # Input: features of both views (C, H, W); pixel grid (H, W, 2) # Output: flow (H, W, 2) at 1/8 resolution def soft_match_flow(feat0, feat1, grid): # Correlate every pixel of view 0 with every pixel of view 1 [eq. 1] corr = flatten(feat0).T @ flatten(feat1) / sqrt(C) # (HW, HW) # Softmax over all targets [eq. 2] prob = softmax(corr, axis=-1) # Expected target coordinate, a weighted average [eq. 3] expected = prob @ flatten(grid) # (HW, 2) # Flow is expected target minus own coordinate [eq. 4] return reshape(expected - flatten(grid), (H, W, 2)) ``` Practical notes for DIFT, from the paper and its code: - **Noise copies.** The default is eight copies. Each copy is a full denoiser pass [arxiv:2306.03881 app. B]. Dropping the averaging costs DIFT_sd 1.6 points per point on SPair-71k (59.5 to 57.9) [arxiv:2306.03881 app. B]. - **Prompts.** For Stable Diffusion, SPair-71k uses "a photo of a [class]" and CUB uses a bird prompt. An empty prompt costs about 2 points on SPair-71k for DIFT_sd (59.5 to 57.6) [arxiv:2306.03881 app. C]. - **Output grid.** Features are at 1/8 or 1/16 of the input size. The code upsamples the feature map to image size before matching, so matches land on an image-size grid while the features stay coarse [arxiv:2306.03881 app. C] [code: eval_spair.py#L94-L95]. - **Keypoint sampling.** Bilinear sampling conventions differ across frameworks. Check the align-corners default before porting the code [code: eval_hpatches.py#L53-L62]. - **Seeds.** Each run draws new noise, so the features change between runs. Fix the seed when outputs must be reproducible [arxiv:2306.03881 app. B]. Practical notes for UniMatch: - **Propagation projections.** The paper's propagation layer has no learned projections [arxiv:2211.05783 §3.3]. The released code applies learned query and key projections before the dot product [code: unimatch/attention.py#L177-L178] [code: unimatch/attention.py#L204-L209]. A code comment says they were kept so the released models need not be retrained. Implementing the paper's version will not reproduce the released weights. - **Stereo variant.** The released stereo scripts use windowed, shifted 1D cross-attention, not the full-width 1D variant of the paper [code: scripts/gmstereo_scale2_regrefine3_train.sh#L28] [code: unimatch/transformer.py#L100-L129]. - **Global matching is quadratic.** A 384 × 512 training crop gives a 48 × 64 grid, so 3,072 positions and about 9.4 M scores per image pair [code: scripts/gmflow_scale2_train.sh#L19] [code: unimatch/matching.py#L15]. The paper does not quantify the memory. - **Flow loss exclusion.** The code drops pixels with ground-truth flow of 400 px or more from the loss [code: loss/flow_loss.py#L14]. The paper does not say this. Practical notes for RoMa: - **Charbonnier scale.** The paper gives c = 0.03 with no coordinate units [arxiv:2305.15404 §3.4]. The training script sets c = 1e-4 on normalised grid coordinates [code: experiments/train_roma_outdoor.py#L219-L220] [code: romatch/losses/robust_loss.py#L89-L92]. Pick one convention and test it on held-out pairs; do not copy one constant over the other. - **Kernel temperature.** The paper gives an inverse temperature of 10 [arxiv:2305.15404 app. A] [app. B]. The code sets T = 0.2 [code: romatch/models/model_zoo/roma_models.py#L153], which is an inverse temperature of 5 in the code's form [code: romatch/models/matcher.py#L187-L205]. - **Coarse memory.** The decoder emits 4,097 logits per coarse pixel. At 560 × 560 the coarse grid is 40 × 40, so each pair carries about 6.6 M coarse logits [arxiv:2305.15404 §4.2] [app. B]. The paper gives no memory figure. - **Training script.** The script warns that it is untested. As written it passes `use_vgg=True` to a constructor that does not take it, so it raises a `TypeError` [code: experiments/train_roma_outdoor.py#L159-L165] [code: romatch/models/encoders.py#L29-L30]. Practical notes for NDF, from the paper and its code: - **Descriptor normalisation.** The paper's descriptor is a plain concatenation [arxiv:2112.05124 eq. 4]. The released code L2-normalises it per query point when features are returned, and the paper does not say whether its numbers used this [code: src/ndf_robot/model/vnn_occupancy_net_pointnet_dgcnn.py#L296-L306]. - **Pose search is under-specified.** The number of random starts, the number of steps and the learning rate are not stated in the paper [arxiv:2112.05124 §II-B]. The released code uses 10 starts and 250 steps by default [code: src/ndf_robot/opt/optimizer.py#L159-L162]. Results depend on these settings, and the search is non-convex. - **Scale unit.** Query-point scale is reported from 0.1 to 10.0, and the unit is not defined [arxiv:2112.05124 tab. III]. The code multiplies inputs by 10 before the network [code: src/ndf_robot/model/vnn_occupancy_net_pointnet_dgcnn.py#L66], so the same scale value means something different there. Practical notes for BEHAVE: - **Loss weights are missing.** The three terms of the fit [arxiv:2204.06950 eq. 1] and the regularisers have no stated weights, so an implementation has to pick them [arxiv:2204.06950 eq. 2]. - **Contact threshold is missing.** The paper uses the same symbol ε for the orientation filter (2 cm), for the network contact test, and for the dataset's 2 cm label. It gives no value for the network contact test [arxiv:2204.06950 §3] [§4.2] [§4.3]. - **Orientation convention.** The nine network outputs are reshaped to a 3 × 3 matrix, but the paper does not say row-major or column-major, and it does not say how a reflection is avoided [arxiv:2204.06950 §4.2]. ## Contradictions Within the DIFT paper, the text and the tables describe the same SPair-71k result in different ways. The introduction says DIFT_sd stays on par with the best supervised method [arxiv:2306.03881 §1]. The per-image table puts it 2.4 points below SCorrSAN [arxiv:2306.03881 tab. 1]. The paper also reports that DIFT_sd beats the best supervised method in 9 of 18 categories [arxiv:2306.03881 §5.2]. The DIFT paper and its code disagree on several points. The paper defines PCK per image as the average of per-image values [arxiv:2306.03881 §5.2]. The code averages one value per image pair [code: eval_spair.py#L85-L127]. The paper's table uses the code's version, so per-pair and per-picture numbers should not be compared. The paper's DAVIS setting uses 28 previous frames [arxiv:2306.03881 app. C], while the code caps the context at 10 frames for one video [code: eval_davis.py#L300-L302]. The UniMatch paper gives two sizes for GMDepth: 7.3 M parameters and 40 ms in one comparison [arxiv:2211.05783 tab. 8], and 4.7 M parameters and 17 ms in another [arxiv:2211.05783 tab. 9]. The paper does not say which configuration each uses. Across the six papers, no benchmark and metric pair is reported by more than one card. RoMa reports pose AUC on ScanNet and UniMatch reports depth on ScanNet, which are different tasks. DragGAN reports landmark distance on face edits, which none of the others measure. NDF reports simulated grasp and place success, and BEHAVE reports body and object vertex error on its own dataset. So no numerical conflict between the papers is known, and their numbers should not be read as a ranking across families. NDF's introduction says overall success is above 85 per cent. That holds for upright objects only; with arbitrary poses the table gives 0.58 to 0.78 [arxiv:2112.05124 §I] [tab. I]. Quote the table, not the headline. ## Papers - [Emergent Correspondence from Image Diffusion](/papers/2306.03881): training-free correspondence from a pretrained diffusion denoiser, with per-task settings and results on SPair-71k, PF-WILLOW, CUB, HPatches and DAVIS. - [Unifying Flow, Stereo and Depth Estimation](/papers/2211.05783): one Transformer matcher with parameter-free softmax matching for optical flow, rectified stereo and posed-image depth, with cross-task transfer results. - [RoMa: Robust Dense Feature Matching](/papers/2305.15404): a coarse-to-fine matcher on frozen DINOv2 features with a Transformer anchor decoder and robust refinement, with results on WxBS, IMC2022 and MegaDepth. - [Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold](/papers/2305.10973): point-based GAN editing whose tracker is a windowed nearest-neighbour search in StyleGAN2 features, with face-landmark and reconstruction results. - [Neural Descriptor Fields: SE(3)-Equivariant Object Representations for Manipulation](/papers/2112.05124): equivariant point descriptors from an occupancy network, used for few-shot pose search, with simulated pick-and-place results. - [BEHAVE: Dataset and Method for Tracking Human Object Interactions](/papers/2204.06950): per-frame tracking of a body and a known object from multi-view depth, with predicted body correspondences and contact, and a non-commercial dataset.