Question about your MTMC ReID pipeline from the LinkedIn post

[FoadG]

Hi Le Duc Dat,

I’m Mohammad Foad Adib, an undergrad computer engineering student from the University of Isfahan, Iran.

I came across your LinkedIn post about the ReID pipeline with Isaac Sim, YOLOv11, ByteTrack, and TransReID. It’s very close to what I’m working on in my research.

I’m trying to understand how different modules like ByteTrack and TransReID affect global ID matching. I’m also hoping to explore what features or layers play a key role, and how accuracy might be improved under different conditions.

I was wondering if it’d be possible to access the implementation you mentioned, or get some guidance on how it’s built. It would be a huge help for my academic project. Of course, I’ll give full credit for anything used or referenced.

Thanks for your time either way.

Mohammad Foad Adib
University of Isfahan
mfoadadib@gmail.com

[LeDat98]

Hi Adib, Thank you for your interest in my multi-camera ReID pipeline! I appreciate your detailed technical questions. First, I need to let you know that I currently cannot share the source code due to project security reasons, as it's being used for commercial products at my company. However, I'm planning to build an open-source version for multi-camera tracking with ReID in the future. Regarding how ByteTrack and TransReID affect global identity consistency: I chose ByteTrack because it's lightweight and easy to deploy. Its main advantage is stable local ID tracking, which significantly helps maintain consistency between local IDs and global IDs (post-ReID) within each session. Technical Architecture: ByteTrack provides local consistency by maintaining temporal stability for local IDs across frames, handling occlusions without losing tracks, and offering low computational overhead for real-time processing. For TransReID integration and global consistency, the pipeline works as: Local ID leads to Feature Extraction, then FAISS Similarity Search, and finally Global ID Assignment. The key strategy is that ByteTrack runs every frame while TransReID processes every 5-10 frames for efficiency. Handling Similar Appearances - Your Core Challenge: Traditional approaches include weighted feature averaging over time, temporal smoothing where new features equal alpha times old features plus one minus alpha times current features, and conflict resolution for multiple local IDs mapping to the same global ID. For advanced spatial-temporal analysis, which I recommend for your research, you should consider homography-based 2D spatial features. You can extract spatial features by converting bbox coordinates to 2D top-view plane using homography matrix transformation of the bottom center foot position. I particularly recommend clustering-based position prediction using three key components: relative position analysis to track position relationships between objects in the same cluster, trajectory pattern learning to analyze movement patterns in 2D space, and spatial-temporal clustering to group objects by movement behavior and relative positions. For enhanced feature fusion, you should combine appearance features from TransReID output, spatial features from 2D position and trajectory data, relative features from position relative to other objects, and temporal features from movement patterns over time. Model Training Recommendations: For your similar appearance research, I suggest a custom dataset strategy where you combine your data with Market-1501 to ensure robust base performance. Focus on collecting similar-looking individuals in various scenarios and include spatial-temporal annotations for trajectory analysis. For the training pipeline, I recommend using a combined dataset of your data plus Market-1501, then take a TransReID model pretrained on Market-1501 and finetune it on the combined dataset with focus on hard negatives. Implementation Framework: You should build a SpatialTemporalReID system that includes an appearance model using TransReID, a spatial analyzer for clustering analysis, and a trajectory predictor. For enhanced matching, the process would be: first extract appearance features using the appearance model, then convert to 2D spatial coordinates by applying homography transformation, analyze relative positions and trajectories using the spatial analyzer, and finally perform combined similarity scoring by weighted combination of appearance and spatial features. This spatial-temporal approach should significantly improve your system's ability to distinguish between similar-looking individuals by leveraging their behavioral and positional patterns in addition to appearance features. Your research focus on similar appearances is fascinating - it's one of the most challenging problems in multi-camera ReID. The combination of appearance and spatial-temporal features should provide much better discrimination capability. Best regards, Le Duc Dat Vào Th 4, 16 thg 7, 2025 vào lúc 04:03 Mohammad Foad < ***@***.***> đã viết:

…

*FoadG* created an issue (LeDat98/CameraTamperingDetection#1) <#1> Hi Le Duc Dat, I’m Mohammad Foad Adib, an undergrad computer engineering student from the University of Isfahan, Iran. I came across your LinkedIn post about the ReID pipeline with Isaac Sim, YOLOv11, ByteTrack, and TransReID. It’s very close to what I’m working on in my research. I’m trying to understand how different modules like ByteTrack and TransReID affect global ID matching. I’m also hoping to explore what features or layers play a key role, and how accuracy might be improved under different conditions. I was wondering if it’d be possible to access the implementation you mentioned, or get some guidance on how it’s built. It would be a huge help for my academic project. Of course, I’ll give full credit for anything used or referenced. Thanks for your time either way. Mohammad Foad Adib University of Isfahan ***@***.*** — Reply to this email directly, view it on GitHub <#1>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A46DMCQSBAOEQL5JZVP2WBD3IVGARAVCNFSM6AAAAACBS2DHYSVHI2DSMVQWIX3LMV43ASLTON2WKOZTGIZTGMZZGQZTMMI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>